SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2212.04356 · 2022

Robust Speech Recognition via Large-Scale Weak Supervision

Alec Radford, Greg Brockman, Ilya Sutskever, Jong Kim, Tao Xu, Christine Mcleavey

arXiv · PDF · Open in the Atlas

Code that ran

49 of 59 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 6 honoured the contract we drafted, 1 violated it and 26 have no contract verdict; 16 failures of our instrument; 10 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2212.04356.json · badge: /api/ran/2212.04356.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
ggerganov/whisper.cpp canonical (classifier label)
classifier basis not recorded on the link
6 ran · 1 no recorded run
briansidp/whisperbiasing community (classifier label)
classifier basis not recorded on the link
10 ran · 3 no recorded run
robflynnyh/long-context-asr community (classifier label)
classifier basis not recorded on the link
10 ran · 0 no recorded run
open-creator/icefall community (classifier label)
classifier basis not recorded on the link
9 ran · 2 no recorded run
kadirnar/whisper-plus community (classifier label)
classifier basis not recorded on the link
4 ran · 3 no recorded run
sanchit-gandhi/whisper-jax community (classifier label)
classifier basis not recorded on the link
4 ran · 0 no recorded run
audioshake/alt-eval community (classifier label)
classifier basis not recorded on the link
3 ran · 0 no recorded run
collabora/whisperlive community (classifier label)
classifier basis not recorded on the link
1 ran · 1 no recorded run
whisperspeech/whisperspeech extension (classifier label)
classifier basis: classifier_2025
1 ran · 0 no recorded run
copy not recorded community
not in the paper's declared repositories
1 ran · 0 no recorded run
FunctionStatusWhere it lives
LinearActivation Ran
ran; our contract draft was wrong, not the code
robflynnyh/long-context-asr/lcasr/components/long_conv.py
code served (permissive licence) · get_code("6ecbcf4f3d00f7ed")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Sequential. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn
import math
from functools import partial

torch.manual_seed(0)

ARGS = {
    'd_input': 4,
    'd_output': 3,
    'bias': True,
    'zero_bias_init': False,
    'transposed': False,
    'initializer': 'normal',
    'activation': 'relu',
    'activate': True,
    'weight_norm': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
blank_p Ran
ran on a synthesized input
robflynnyh/long-context-asr/exp/train_utterance_meta_loss_pred.py
code served (permissive licence) · get_code("3787ac531561b6c6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from transformers import AutoTokenizer

torch.manual_seed(0)

# Create a minimal tokenizer mock that has a decode method
class MockTokenizer:
    def decode(self, token_ids):
        return "sample decoded text"

# Create logits tensor: shape (batch=2, sequence=3, vocab=5)
# vocab_size=5, where the last token (index 4) is the blank token
logits = torch.randn(2, 3, 5)

# Create a tokenizer instance
tokenizer = MockTokenizer()

ARGS = {
    "logits": logits,
    "tokenizer": tokenizer,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
blank_p Ran
ran on a synthesized input
robflynnyh/long-context-asr/exp/train_utterance_meta_loss_pred_ce.py
code served (permissive licence) · get_code("9f16788e5da13dfa")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from transformers import AutoTokenizer

torch.manual_seed(0)

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def decode(self, token_ids):
        return "".join([f"token_{i}" for i in token_ids])

# Create logits tensor: (batch=2, seq_len=3, vocab_size=5)
# The function uses logits.shape[-1] to get vocab size
# It treats the last token index (vocab_size - 1) as a "blank" token
logits = torch.randn(2, 3, 5)

# Create tokenizer
tokenizer = MinimalTokenizer()

ARGS = {
    "logits": logits,
    "tokenizer": tokenizer,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
byte_decode Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/byte_utils.py
code served (permissive licence) · get_code("59d410aa19bdef96")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# Define the constants and helper dicts from the function definition
PRINTABLE_BASE_CHARS = [256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 321, 322, 323, 324, 325, 326, 327, 328, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422]

BYTE_TO_BCHAR = {b: chr(PRINTABLE_BASE_CHARS[b]) for b in range(256)}

BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()}

# Create a valid test string by encoding some UTF-8 bytes through the byte_decode mapping
# The function takes a string 'x' and converts it via BCHAR_TO_BYTE to bytes, then decodes as UTF-8
test_string = chr(PRINTABLE_BASE_CHARS[72]) + chr(PRINTABLE_BASE_CHARS[101]) + chr(PRINTABLE_BASE_CHARS[108]) + chr(PRINTABLE_BASE_CHARS[108]) + chr(PRINTABLE_BASE_CHARS[111])  # "Hello"

ARGS = {
    'x': test_string
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
byte_encode Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/byte_utils.py
code served (permissive licence) · get_code("004eb0067bd43754")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

WHITESPACE_NORMALIZER = re.compile('\\s+')
SPACE = chr(32)

PRINTABLE_BASE_CHARS = [256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 321, 322, 323, 324, 325, 326, 327, 328, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422]

BYTE_TO_BCHAR = {b: chr(PRINTABLE_BASE_CHARS[b]) for b in range(256)}

def byte_encode(x: str) -> str:
    normalized = WHITESPACE_NORMALIZER.sub(SPACE, x)
    return ''.join([BYTE_TO_BCHAR[b] for b in normalized.encode('utf-8')])

ARGS = {
    'x': 'hello  world'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
check_if_already_evaluated Ran
ran on a synthesized input
robflynnyh/long-context-asr/eval/eval_manager.py
code served (permissive licence) · get_code("9140147b0fd2c440")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
import pandas as pd
from torch import nn

torch.manual_seed(0)

# Create a minimal DataFrame that matches the filtering operations in the function
data = {
    'checkpoint': ['/path/to/model.pt', '/path/to/model.pt', '/path/to/other.pt'],
    'dataset': ['dataset_a', 'dataset_a', 'dataset_b'],
    'split': ['train', 'val', 'train'],
    'seq_len': [128, 128, 256],
    'overlap_ratio': [0.5, 0.5, 0.3]
}
cur_df = pd.DataFrame(data)

# Create a simple model (not actually used in the function body)
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

model = SimpleModel()

# Create args object with required attributes
class Args:
    def __init__(self):
        self.seq_len = 128
        self.overlap = 0.5

args = Args()

# Create the ARGS dictionary
ARGS = {
    'model_save_path': '/path/to/model.pt',
    'cur_df': cur_df,
    'dataset': 'dataset_a',
    'split': 'train',
    'args': args
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
check_in_utt Ran
ran on a synthesized input
briansidp/whisperbiasing/dataloader.py
pointer only (licence: MIT) · get_code("fa47326ca5e0886b")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

collate_wrapper Ran
ran on a synthesized input
briansidp/whisperbiasing/dataloader.py
pointer only (licence: MIT) · get_code("c4c3b54ad53b1fce")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

compression_ratio Ran
ran on a synthesized input
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/decoding.py
code served (permissive licence) · get_code("5632d6609129dfa3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import zlib

# The function takes a single parameter 'text' which should be a string
# It encodes the string to UTF-8 bytes, compresses it, and calculates a ratio
# We need to provide a valid string argument

ARGS = {
    "text": "The quick brown fox jumps over the lazy dog. " * 10
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_alignments_to_tensor Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/ali.py
code served (permissive licence) · get_code("8a64b592c82c0d10")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Construct alignments dict with utterance IDs as keys and lists of ints as values
alignments = {
    "utt_001": [0, 1, 2, 1, 0],
    "utt_002": [3, 2, 1, 0, 1, 2],
    "utt_003": [0, 0, 1],
}

# Device
device = torch.device("cpu")

ARGS = {
    "alignments": alignments,
    "device": device,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
correct_for_bias_scale_order_inversion Ran
ran on a synthesized input
ggerganov/whisper.cpp/models/convert-whisper-to-coreml.py
code served (permissive licence) · get_code("b41e987fee8a92ff")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Construct the state_dict with bias and weight tensors
state_dict = {
    'layer.bias': torch.tensor([1.0, 2.0, 3.0, 4.0]),
    'layer.weight': torch.tensor([2.0, 2.0, 2.0, 2.0])
}

# Define prefix to match the keys in state_dict
prefix = 'layer.'

# Local metadata - typically empty dict for loading state dicts
local_metadata = {}

# Strict mode flag
strict = True

# Missing keys list
missing_keys = []

# Unexpected keys list
unexpected_keys = []

# Error messages list
error_msgs = []

ARGS = {
    'state_dict': state_dict,
    'prefix': prefix,
    'local_metadata': local_metadata,
    'strict': strict,
    'missing_keys': missing_keys,
    'unexpected_keys': unexpected_keys,
    'error_msgs': error_msgs
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
dot_product_attention Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
sanchit-gandhi/whisper-jax/whisper_jax/layers.py
pointer only (licence: Apache-2.0) · get_code("2000f34f95c683d6")
How it ran

The body of this sample is withheld: Apache-2.0, as GitHub detects it on this paper's repository, sanchit-gandhi/whisper-jax, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

dtw_cpu Ran
ran, honoured the contract we drafted
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/timing.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from roudimit/whisper-flamingo, is pointer only (licence: NOASSERTION) · get_code("76330ed965e60af5")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from roudimit/whisper-flamingo, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

exact_div Ran
ran, honoured the contract we drafted
briansidp/whisperbiasing/whisper/utils.py
pointer only (licence: MIT); the stored copy get_code serves, from qwenlm/qwen-audio, is pointer only (licence: NOASSERTION) · get_code("6e0808fc4828513a")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

find_checkpoints Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/checkpoint.py
code served (permissive licence) · get_code("87c9b24deaf27799")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os
from pathlib import Path

# Create a temporary directory with some checkpoint files
temp_dir = tempfile.mkdtemp()

# Create some dummy checkpoint files for testing
checkpoint_files = [
    "checkpoint-1.pt",
    "checkpoint-20.pt",
    "checkpoint-300.pt",
    "checkpoint-4000.pt"
]

for checkpoint_file in checkpoint_files:
    filepath = os.path.join(temp_dir, checkpoint_file)
    with open(filepath, 'w') as f:
        f.write("dummy checkpoint")

# Define arguments for the function
ARGS = {
    "out_dir": Path(temp_dir),
    "iteration": 0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
format_time Ran
ran on a synthesized input
collabora/whisperlive/whisper_live/utils.py
code served (permissive licence) · get_code("9649e15864c56627")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "s": 3661.5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_args Ran
ran on a synthesized input
robflynnyh/long-context-asr/eval/eval_manager.py
code served (permissive licence) · get_code("63302892207b7f5b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ArgsClass. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.path = "/mock/model/path"
        self.seq_len = 128
        self.overlap_ratio = 0.875
        self._args = {'model_arg1': 'value1', 'model_arg2': 42}
    
    def get(self, key, default=None):
        if key == 'overlap_ratio':
            return self.overlap_ratio
        elif key == 'args':
            return self._args
        return default

class MockConfig(dict):
    def __init__(self, name=None, args=None):
        super().__init__()
        self.name = name or "default_dataset"
        self._args = args or {'config_arg1': 'value2'}
    
    def get(self, key, default=None):
        if key == 'args':
            return self._args
        return super().get(key, default)

model = MockModel()
config = MockConfig(name="test_config", args={'config_arg1': 'value2', 'config_arg2': 100})
split = "train"
dataset_config = MockConfig(name="dataset_name", args={'dataset_arg1': 'value3', 'dataset_arg2': 0.5})

ARGS = {
    'config': config,
    'split': split,
    'model': model,
    'dataset_config': dataset_config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArgsClass — a ArgsClass output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ArgsClass output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_class_name Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/diagnostics.py
code served (permissive licence) · get_code("b865558252e1566a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

# Test with a basic nn.Module
class SimpleModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Test with Balancer module
class Balancer(nn.Module):
    def __init__(self):
        super().__init__()
        self.min_positive = 0.1
        self.max_positive = 0.9
        self.min_abs = 0.01
        self.max_abs = 0.99
    
    def forward(self, x):
        return x

# Test with ActivationBalancer module
class ActivationBalancer(nn.Module):
    def __init__(self):
        super().__init__()
        self.min_positive = 0.2
        self.max_positive = 0.8
        self.min_abs = 0.05
        self.max_abs = 0.95
    
    def forward(self, x):
        return x

# Test with AbsValuePenalizer module
class AbsValuePenalizer(nn.Module):
    def __init__(self):
        super().__init__()
        self.limit = 1.5
    
    def forward(self, x):
        return x

torch.manual_seed(0)

# Create instances to test with
simple_module = SimpleModule()
balancer_module = Balancer()
activation_balancer_module = ActivationBalancer()
abs_value_penalizer_module = AbsValuePenalizer()

ARGS = {
    'module': simple_module
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_data_to_save Ran
ran on a synthesized input
robflynnyh/long-context-asr/eval/eval_manager.py
code served (permissive licence) · get_code("ae28fa768d63729b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockConfig:
    def __init__(self):
        self.args = MockArgs()

class MockArgs:
    def __init__(self):
        self.model_class = "TestModel"

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.name = "test_model"
        self.path = "/path/to/checkpoint.pt"
        self.repeat = 1
        self.seq_len = 512
        self.overlap_ratio = 0.5

    def forward(self, x):
        return x

config = MockConfig()

wers = [
    {
        'wer': 0.15,
        'recording': 'rec_001',
        'words': 100,
        'ins_rate': 0.02,
        'del_rate': 0.05,
        'sub_rate': 0.08
    },
    {
        'wer': 0.12,
        'recording': 'rec_002',
        'words': 150,
        'ins_rate': 0.01,
        'del_rate': 0.04,
        'sub_rate': 0.07
    }
]

split = "test"
dataset = "librispeech"
model = MockModel()

ARGS = {
    "config": config,
    "wers": wers,
    "split": split,
    "dataset": dataset,
    "model": model
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_dtype Ran
ran on a synthesized input
robflynnyh/long-context-asr/exp/train_utterance_meta_loss_pred.py
code served (permissive licence) · get_code("ad63708b6da385d5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dtype. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# The function get_dtype takes a single parameter 'dtype' which is a string
# It maps the string to the corresponding torch.dtype
# We need to provide valid test cases for this function

torch.manual_seed(0)

# Test with a valid dtype string
ARGS = {
    'dtype': 'float32'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dtype — a dtype output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dtype output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_initializer Ran
ran on a synthesized input
robflynnyh/long-context-asr/lcasr/components/long_conv.py
code served (permissive licence) · get_code("ee3c8d2a605ac8bc")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned partial. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from functools import partial

torch.manual_seed(0)

# Test with different initializer names and activations
ARGS = {
    'name': 'uniform',
    'activation': 'relu'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type partial — a partial output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a partial output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_short_name Ran
ran; our contract draft was wrong, not the code
ggerganov/whisper.cpp/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
pointer only (licence: MIT); the stored copy get_code serves, from eziyoo/LLMs-on-Devices, is pointer only (licence: NONE) · get_code("a5b8c22a3ac7b238")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from eziyoo/LLMs-on-Devices, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_tensor_stats Ran
ran on a synthesized input
open-creator/icefall/icefall/diagnostics.py
code served (permissive licence) · get_code("d1162925e2d53784")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create test tensors with distinct dimensions
# Using shape (2, 3, 4, 5) to have distinguishable dimensions
x = torch.randn(2, 3, 4, 5)

# Test with different dims and stats_types
dim = 1
stats_type = "abs"

ARGS = {
    "x": x,
    "dim": dim,
    "stats_type": stats_type,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
hz_to_mel Ran
ran on a synthesized input
ggerganov/whisper.cpp/models/convert-parakeet-to-ggml.py
code served (permissive licence) · get_code("9720e570a35f6d55")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

def hz_to_mel(freq):
    return 2595.0 * np.log10(1.0 + freq / 700.0)

# The function takes a frequency value (or array of frequencies)
# and converts it to mel scale using the formula provided.
# freq can be a scalar, list, or numpy array

ARGS = {
    "freq": np.array([440.0, 880.0, 1760.0, 3520.0])
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
laplace Ran
ran; our contract draft was wrong, not the code
robflynnyh/long-context-asr/lcasr/components/long_conv.py
code served (permissive licence) · get_code("a799a46446fafd99")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import math

torch.manual_seed(0)

# Create a small tensor for x parameter
x = torch.randn(2, 4)

# Default parameters are already specified in the function signature
mu = 0.707107
sigma = 0.282095

ARGS = {
    "x": x,
    "mu": mu,
    "sigma": sigma,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_alignments Ran
ran on a synthesized input
open-creator/icefall/icefall/ali.py
code served (permissive licence) · get_code("a972a5df2771f7a5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import tempfile
import os

torch.manual_seed(0)

# Create a temporary file with alignment data
temp_dir = tempfile.mkdtemp()
temp_filename = os.path.join(temp_dir, "alignments.pt")

# Create alignment data structure matching what load_alignments expects
ali_dict = {
    'subsampling_factor': 4,
    'alignments': {
        'utt_1': [0, 1, 2, 3, 4, 5],
        'utt_2': [1, 2, 3, 4, 5, 6, 7],
        'utt_3': [0, 0, 1, 2, 3]
    }
}

# Save the alignment dictionary
torch.save(ali_dict, temp_filename)

ARGS = {
    'filename': temp_filename
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_model_config Ran
ran on a synthesized input
ggerganov/whisper.cpp/models/convert-parakeet-to-ggml.py
code served (permissive licence) · get_code("ba1be9f251ecd80a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import yaml
import tempfile
import os

# Create a temporary YAML config file
temp_dir = tempfile.mkdtemp()
config_path = os.path.join(temp_dir, "test_config.yaml")

# Write a simple config to the temporary file
config_data = {
    "model": "transformer",
    "hidden_size": 768,
    "num_layers": 12,
    "vocab_size": 30522,
    "learning_rate": 0.001,
    "batch_size": 32
}

with open(config_path, 'w', encoding='utf-8') as f:
    yaml.dump(config_data, f)

ARGS = {
    "config_path": config_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
yaml — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
logits_to_probs Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
whisperspeech/whisperspeech/whisperspeech/inference.py
code served (permissive licence) · get_code("8a8edc42e5e25cea")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family float, shape (2, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small logits tensor
# Shape: (batch_size=2, vocab_size=5)
logits = torch.randn(2, 5)

# Temperature parameter
T = 1.0

# Top-k parameter (will keep top 3 logits)
top_k = 3

ARGS = {
    'logits': logits,
    'T': T,
    'top_k': top_k
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 5) — shape (2, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
lookup_alignments Ran
ran, honoured the contract we drafted
open-creator/icefall/icefall/ali.py
code served (permissive licence) · get_code("a1f5678203ff47e8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 4, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create small alignment tensors for two utterances
# First utterance: alignment [1, 3, 2] (3 frames)
# Second utterance: alignment [1, 0, 4, 2] (4 frames)
ali_1 = torch.tensor([1, 3, 2], dtype=torch.long)
ali_2 = torch.tensor([1, 0, 4, 2], dtype=torch.long)

# Create the alignments dictionary
alignments = {
    "utt_001": ali_1,
    "utt_002": ali_2,
}

# Create the cut_ids list
cut_ids = ["utt_001", "utt_002"]

# Number of classes (max token ID + 1)
num_classes = 5

# Log score for non-alignment positions
log_score = -10.0

ARGS = {
    "cut_ids": cut_ids,
    "alignments": alignments,
    "num_classes": num_classes,
    "log_score": log_score,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 4, 5) — shape (2, 4, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_lexical_tree Ran
ran; our contract draft was wrong, not the code
briansidp/whisperbiasing/dataloader.py
pointer only (licence: MIT) · get_code("c45f3b632ec36534")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

median_filter Ran
ran on a synthesized input
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/timing.py
code served (permissive licence) · get_code("6f669ba80cbbf823")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 3, dtype family float, shape (2, 3, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from scipy import signal

# Set random seed for reproducibility
np.random.seed(0)

# Create a small test array with shape (batch, height, width)
# Using distinct dimensions: 2, 3, 5
x = np.random.randn(2, 3, 5).astype(np.float32)

# Filter width must be a positive odd number
filter_width = 3

ARGS = {
    "x": x,
    "filter_width": filter_width,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family float, shape (2, 3, 5) — shape (2, 3, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
mel_to_hz Ran
ran on a synthesized input
ggerganov/whisper.cpp/models/convert-parakeet-to-ggml.py
code served (permissive licence) · get_code("40e421c01b760f1c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# The function mel_to_hz takes a mel value (scalar or array-like)
# and converts it to Hz using a formula
# It performs arithmetic operations: division, power, subtraction, multiplication

# Create a simple mel value to convert
mel_value = 2595.0

ARGS = {
    "mel": mel_value
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
multinomial_sample_one_no_sync Ran
ran; our contract draft was wrong, not the code
this paper's copy was not recorded; identical code first harvested from Cornell-RelaxML/qtip
pointer only; the stored copy get_code serves, from Cornell-RelaxML/qtip, is pointer only (licence: GPL-3.0) · get_code("0b5f90e8856ad7af")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

normalize_lyrics Ran
ran on a synthesized input
audioshake/alt-eval/src/alt_eval/normalization.py
code served (permissive licence) · get_code("c4ee356de1bd8c77")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import regex as re

def normalize_lyrics(text: str) -> str:
    """Normalize lyrics to improve adherence to the Jam-ALT annotation guide.

    Unwanted end-of-line punctuation is removed and the first letter of each line is uppercased.

    This is a relatively crude normalization method that should generally improve the results at
    least for languages that use the Latin script. However, it may make the results worse if the
    line break predictions are not accurate.
    """
    text = re.sub('(?<!^)[^\\w\\n!?\'´''\\"""»)]+ *$', '', text, flags=re.MULTILINE)
    text = re.sub('^([^\\w\\n]*\\w)', lambda m: m.group(1).upper(), text, flags=re.MULTILINE)
    return text


ARGS = {
    "text": "hello world\nthis is a test...\nanother line!"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
regex — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
optional_int Ran
ran, honoured the contract we drafted
briansidp/whisperbiasing/whisper/utils.py
pointer only (licence: MIT) · get_code("e56996b20b9c19cb")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pad_or_trim Ran
ran; our contract draft was wrong, not the code
briansidp/whisperbiasing/whisper/audio.py
pointer only (licence: MIT) · get_code("9f13de1c23e4003d")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pjit Ran
ran on a synthesized input
sanchit-gandhi/whisper-jax/whisper_jax/partitioner.py
pointer only (licence: Apache-2.0) · get_code("87bdf3970749dc11")
How it ran

The body of this sample is withheld: Apache-2.0, as GitHub detects it on this paper's repository, sanchit-gandhi/whisper-jax, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pjit_with_cpu_fallback Ran
ran on a synthesized input
sanchit-gandhi/whisper-jax/whisper_jax/partitioner.py
pointer only (licence: Apache-2.0) · get_code("d35e5ea9b14ce05e")
How it ran

The body of this sample is withheld: Apache-2.0, as GitHub detects it on this paper's repository, sanchit-gandhi/whisper-jax, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

remove_symbols Ran
ran; our contract draft was wrong, not the code
briansidp/whisperbiasing/whisper/normalizers/basic.py
pointer only (licence: MIT) · get_code("2206d84b6900ab3e")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

remove_symbols_and_diacritics Ran
ran; our contract draft was wrong, not the code
briansidp/whisperbiasing/whisper/normalizers/basic.py
pointer only (licence: MIT) · get_code("cef27f679a755971")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

rename_keys Ran
ran on a synthesized input
ggerganov/whisper.cpp/models/convert-h5-to-coreml.py
code served (permissive licence) · get_code("e65744d3337badd5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a sample state dict with keys that match the WHISPER_MAPPING patterns
s_dict = {
    'model.layers.0.fc1.weight': torch.randn(4, 3),
    'model.layers.0.fc2.weight': torch.randn(3, 4),
    'model.layers.0.final_layer_norm.weight': torch.randn(4),
    'model.layers.0.self_attn.q_proj.weight': torch.randn(4, 3),
    'model.layers.0.self_attn.k_proj.weight': torch.randn(4, 3),
    'model.layers.0.self_attn.v_proj.weight': torch.randn(4, 3),
    'model.layers.0.self_attn_layer_norm.weight': torch.randn(4),
    'model.layers.0.self_attn.out_proj.weight': torch.randn(3, 4),
    'model.layers.0.encoder_attn.q_proj.weight': torch.randn(4, 3),
    'model.layers.0.encoder_attn.k_proj.weight': torch.randn(4, 3),
    'model.layers.0.encoder_attn.v_proj.weight': torch.randn(4, 3),
    'model.layers.0.encoder_attn_layer_norm.weight': torch.randn(4),
    'model.layers.0.encoder_attn.out_proj.weight': torch.randn(3, 4),
    'decoder.layer_norm.weight': torch.randn(4),
    'encoder.layer_norm.weight': torch.randn(4),
    'embed_tokens.weight': torch.randn(10, 4),
    'encoder.embed_positions.weight': torch.randn(5, 4),
    'decoder.embed_positions.weight': torch.randn(5, 4),
    'layer_norm.weight': torch.randn(4),
}

ARGS = {
    's_dict': s_dict
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
rgetattr Ran
ran on a synthesized input
robflynnyh/long-context-asr/exp/run_launcher.py
code served (permissive licence) · get_code("bdbf18d0ee9006f7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer1 = nn.Linear(4, 3)
        self.layer2 = nn.Linear(3, 2)
        self.nested_attr = "test_value"
    
    def forward(self, x):
        return self.layer2(self.layer1(x))

obj = SimpleModel()
attr = "layer1.in_features"

ARGS = {
    "obj": obj,
    "attr": attr
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sinusoids Ran
ran, honoured the contract we drafted
briansidp/whisperbiasing/whisper/model.py
pointer only (licence: MIT) · get_code("e529c9641c178fe0")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

sinusoids Ran
ran, honoured the contract we drafted
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/torch_whisper.py
code served (permissive licence) · get_code("407b4297398ff29d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (8, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import torch

torch.manual_seed(0)

ARGS = {
    "length": 8,
    "channels": 4,
    "max_timescale": 10000,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (8, 4) — shape (8, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
smart_byte_decode Ran
ran; our contract draft was wrong, not the code
open-creator/icefall/icefall/byte_utils.py
code served (permissive licence) · get_code("226350299950338b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# The function smart_byte_decode takes a single parameter: a string x
# It processes the string through byte decoding logic

# Create a simple test input string
# The function expects a string that can be decomposed into byte characters
# that map to the PRINTABLE_BASE_CHARS encoding

PRINTABLE_BASE_CHARS = [256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 308, 309, 310, 311, 312, 313, 314, 315, 316, 317, 318, 321, 322, 323, 324, 325, 326, 327, 328, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422]

BYTE_TO_BCHAR = {b: chr(PRINTABLE_BASE_CHARS[b]) for b in range(256)}

BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()}

# Create a valid input string using the encoding
# Convert "hello" to bytes, then to bchars
test_bytes = b"hello"
test_bchar_string = "".join([BYTE_TO_BCHAR[b] for b in test_bytes])

ARGS = {
    "x": test_bchar_string
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
str2bool Ran
ran, violated the contract we drafted
briansidp/whisperbiasing/whisper/utils.py
pointer only (licence: MIT) · get_code("176b6ba6c2c2940c")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

to_rich_tokens Ran
ran on a synthesized input
audioshake/alt-eval/src/alt_eval/tokenizer.py
code served (permissive licence) · get_code("2ef9c9c7ee284aaf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from dataclasses import dataclass, field
import regex as re

@dataclass
class Token:
    """A "rich" token (with a set of associated tags)."""
    text: str
    tags: set = field(default_factory=set)

WORD = 'word'
PUNCT = 'punctuation'
LINE = 'line_break'
SECT = 'section_break'
PAREN = 'parenthesis'

def to_rich_tokens(tokens: list[str]) -> list[Token]:
    """Convert a list of plain (string) tokens to a list of rich tokens."""
    result: list[Token] = []
    for token in tokens:
        rich_token = Token(text=token)
        result.append(rich_token)
        if re.search('\\w', token):
            rich_token.tags.add(WORD)
        elif token == '\n':
            rich_token.text = '<L>'
            rich_token.tags.add(LINE)
        elif '\n' in token:
            assert token == '\n\n'
            rich_token.text = '<S>'
            rich_token.tags.add(SECT)
        elif token in '()':
            rich_token.tags.add(PAREN)
        else:
            rich_token.tags.add(PUNCT)
            if rich_token.text == '@-@':
                rich_token.text = '-'
    return result

# Test tokens covering various cases
tokens = [
    'hello',           # word (contains \w)
    'world',           # word (contains \w)
    '.',               # punctuation
    ',',               # punctuation
    '(',               # parenthesis
    ')',               # parenthesis
    '\n',              # line break
    '\n\n',            # section break
    '@-@',             # punctuation that gets transformed
]

ARGS = {
    'tokens': tokens
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, regex, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
tokens_as_words Ran
ran on a synthesized input
audioshake/alt-eval/src/alt_eval/tokenizer.py
code served (permissive licence) · get_code("3d1de040cc006f6a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import copy
from dataclasses import dataclass, field
import regex as re

@dataclass
class Token:
    """A "rich" token (with a set of associated tags)."""
    text: str
    tags: set = field(default_factory=set)

WORD = 'word'

def tokens_as_words(tokens: list[Token]) -> list[Token]:
    """Process a list of rich tokens to filter out any non-word characters."""
    result: list[Token] = []
    for token in tokens:
        if WORD in token.tags:
            token = copy.deepcopy(token)
            token.text = re.sub("[^\\w']", '', token.text)
            result.append(token)
    return result

# Construct test tokens with varying tags
token1 = Token(text="hello!", tags={WORD, "noun"})
token2 = Token(text="world@123", tags={WORD})
token3 = Token(text="don't", tags={WORD})
token4 = Token(text="###", tags={"punctuation"})  # No WORD tag, should be filtered out
token5 = Token(text="test-case", tags={WORD, "verb"})

tokens_list = [token1, token2, token3, token4, token5]

ARGS = {
    "tokens": tokens_list
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
regex — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
with_sharding_constraint Ran
ran on a synthesized input
sanchit-gandhi/whisper-jax/whisper_jax/partitioner.py
pointer only (licence: Apache-2.0) · get_code("e5a9a9806511efb5")
How it ran

The body of this sample is withheld: Apache-2.0, as GitHub detects it on this paper's repository, sanchit-gandhi/whisper-jax, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

attach_diagnostics Not yet run
unverified (harvested, no recorded run)
open-creator/icefall/icefall/diagnostics.py
code served (permissive licence) · get_code("5969d689c8a84ef2")
How it ran

no invocation recorded.

average_checkpoints Not yet run
unverified (harvested, no recorded run)
open-creator/icefall/icefall/checkpoint.py
code served (permissive licence) · get_code("69124b03f4b749bd")
How it ran

no invocation recorded.

backtrace Not yet run
unverified (harvested, no recorded run)
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/timing.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from roudimit/whisper-flamingo, is pointer only (licence: NOASSERTION) · get_code("41c654294ba9ec56")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from roudimit/whisper-flamingo, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

build_tokenizer Not yet run
unverified (harvested, no recorded run)
briansidp/whisperbiasing/whisper/tokenizer.py
pointer only (licence: MIT) · get_code("acb6a55d43c7f465")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

generate Not yet run
unverified (harvested, no recorded run)
ggerganov/whisper.cpp/models/generate-parakeet-test-model.py
code served (permissive licence) · get_code("fef7828e835c938d")
How it ran

no invocation recorded.

get_encoding Not yet run
unverified (harvested, no recorded run)
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/tokenizer.py
code served (permissive licence) · get_code("b3916c1b2edcf996")
How it ran

no invocation recorded.

get_tokenizer Not yet run
unverified (harvested, no recorded run)
kadirnar/whisper-plus/whisperplus/pipelines/lightning_whisper_mlx/tokenizer.py
code served (permissive licence) · get_code("f2f52f6a90834243")
How it ran

no invocation recorded.

load_audio Not yet run
unverified (harvested, no recorded run)
collabora/whisperlive/whisper_live/diarization.py
code served (permissive licence) · get_code("34dd5baae0e5830e")
How it ran

no invocation recorded.

mel_filters Not yet run
unverified (harvested, no recorded run)
briansidp/whisperbiasing/whisper/audio.py
pointer only (licence: MIT); the stored copy get_code serves, from YasserdahouML/Lip2Vec, is pointer only (licence: NONE) · get_code("edf924293b4b2375")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

transcribe Not yet run
unverified (harvested, no recorded run)
briansidp/whisperbiasing/whisper/transcribe.py
pointer only (licence: MIT) · get_code("738c6e74fd16381d")
How it ran

The body of this sample is withheld: MIT, as GitHub detects it on this paper's repository, briansidp/whisperbiasing, is in the permissive set, but a licence decision keeps the body a pointer: the reason for that decision is not recorded with it. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 59 samples lifted from this paper's linked repositories, 59 carry a function identity and 0 do not. Of the 57 distinct functions with an identity, 28 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 0 the same month, 26 later) and 29 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
sinusoids 9 other papers
9 later
2306.05284 (2023-06)
Simple and Controllable Music Generation
2307.03183 (2023-07)
Whisper-AT: Noise-Robust Automatic Speech Recognizers are Also Strong General Audio Event Taggers
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
and 6 more not named here
multinomial_sample_one_no_sync 8 other papers
8 later
2306.05284 (2023-06)
Simple and Controllable Music Generation
2308.02560 (2023-08)
From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion
2402.04396 (2024-02)
QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks
and 5 more not named here
exact_div 6 other papers
6 later
2306.01428 (2023-06)
Improved DeepFake Detection Using Whisper Features
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
and 3 more not named here
dot_product_attention 5 other papers
3 dated earlier, 2 later
2106.03598 (2021-06)
SciFive: a text-to-text transformer model for biomedical literature
2112.00791 (2021-12)
Controlling Conditional Language Models without Catastrophic Forgetting
2206.05408 (2022-06)
Multi-instrument Music Synthesis with Spectrogram Diffusion
and 2 more not named here
pad_or_trim 5 other papers
5 later
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
2506.08967 (2025-06)
Step-Audio-AQAA: a Fully End-to-End Expressive Large Audio Language Model
and 2 more not named here
optional_int 4 other papers
4 later
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
and 1 more not named here
str2bool 4 other papers
4 later
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
and 1 more not named here
laplace 3 other papers
1 dated earlier, 2 later · 2 only as an assembly brick
2110.13985 (2021-10)
Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers
2306.15794 (2023-06)
HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution
2403.03234 (2024-03)
Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling
LinearActivation 2 other papers
1 dated earlier, 1 later
2210.09298 (2022-10)
What Makes Convolutional Models Great on Long Sequence Modeling?
2608.24597 (2026-08)
Taming foundation model with invariance-oriented pre-training for broad-spectrum EEG analysis across signal-level, brain-state, and brain-health tasks
find_checkpoints 2 other papers
2 later
2506.14434 (2025-06)
Unifying Streaming and Non-streaming Zipformer-based ASR
2512.23278 (2025-12)
Flow2GAN: Hybrid Flow Matching and GAN with Multi-Resolution Network for Few-step High-Fidelity Audio Generation
get_class_name 2 other papers
2 later
2506.14434 (2025-06)
Unifying Streaming and Non-streaming Zipformer-based ASR
2512.23278 (2025-12)
Flow2GAN: Hybrid Flow Matching and GAN with Multi-Resolution Network for Few-step High-Fidelity Audio Generation
mel_filters 2 other papers
2 later
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2506.08967 (2025-06)
Step-Audio-AQAA: a Fully End-to-End Expressive Large Audio Language Model
attach_diagnostics 1 other paper
1 later
2512.23278 (2025-12)
Flow2GAN: Hybrid Flow Matching and GAN with Multi-Resolution Network for Few-step High-Fidelity Audio Generation
backtrace 1 other paper
1 later
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
build_tokenizer 1 other paper
1 later
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
byte_decode 1 other paper
1 later
2506.14434 (2025-06)
Unifying Streaming and Non-streaming Zipformer-based ASR
byte_encode 1 other paper
1 later
2506.14434 (2025-06)
Unifying Streaming and Non-streaming Zipformer-based ASR
convert_alignments_to_tensor 1 other paper
1 later
2506.14434 (2025-06)
Unifying Streaming and Non-streaming Zipformer-based ASR
correct_for_bias_scale_order_inversion 1 other paper
1 dated earlier
1701.07204 (2017-01)
Fast Exact k-Means, k-Medians and Bregman Divergence Clustering in 1D
dtw_cpu 1 other paper
1 later
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation

Shown: the 20 functions present in the most other papers. 8 more shared functions are not listed here.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

We study the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results but in a zeroshot transfer setting without the need for any finetuning. When compared to humans, the models approach their accuracy and robustness. We are releasing models and inference code to serve as a foundation for further work on robust speech processing.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2212.04356")
get_code_for_paper("2212.04356")
have("2212.04356")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2212.04356.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2212.04356.svg)](https://syntology.ai/paper/2212.04356)

Connect an agent — have() is free.