SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2204.07705 · EMNLP · 2022

SUPER-NATURALINSTRUCTIONS: Generalization via Declarative Instructions on 1600+ NLP Tasks

Yejin Choi, Swaroop Mishra, Anjana Arunkumar, Chitta Baral, Hannaneh Hajishirzi, Yizhong Wang, Atharva Naik, Xudong Shen, Siddhartha Mishra, Daniel Khashabi, Arjun Ashok, Pulkit Verma, and 28 more

arXiv · PDF · Open in the Atlas

Code that ran

17 of 28 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 13 have no contract verdict; 3 failures of our instrument; 11 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/2204.07705.json · badge: /api/ran/2204.07705.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
naver-ai/rethinking-proxy-reward community (classifier label)
classifier basis not recorded on the link
7 ran · 1 no recorded run
renzelou/pick-rank community (classifier label)
classifier basis not recorded on the link
6 ran · 6 no recorded run
hyintell/citb community
Papers with Code archive
2 ran · 2 no recorded run
zhaochenyang20/Prompt2Model-Self-Guide community
Papers with Code archive
1 ran · 0 no recorded run
yizhongw/tk-instruct community (classifier label)
classifier basis not recorded on the link
1 ran · 0 no recorded run
allenai/natural-instructions-expansion community
Papers with Code archive
0 ran · 2 no recorded run
FunctionStatusWhere it lives
Adapter Ran
ran on a synthesized input
hyintell/citb/continual_learning/AdapterT5.py
pointer only (licence: NONE) · get_code("d4d90150a63de630")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hyintell/citb 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.

FindAllSuffix Ran
ran on a synthesized input
renzelou/pick-rank/data_proess.py
code served (permissive licence) · get_code("edf1add18a07a762")
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 os
import tempfile
import shutil

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

# Create some test files with 'json' suffix
test_file_1 = os.path.join(temp_dir, 'data.json')
test_file_2 = os.path.join(temp_dir, 'config.json')
test_file_3 = os.path.join(temp_dir, 'info.txt')
test_file_4 = os.path.join(temp_dir, 'results.json')

# Write dummy content to these files
with open(test_file_1, 'w') as f:
    f.write('{}')
with open(test_file_2, 'w') as f:
    f.write('{}')
with open(test_file_3, 'w') as f:
    f.write('test')
with open(test_file_4, 'w') as f:
    f.write('{}')

# Create a subdirectory to test that directories are skipped
subdir = os.path.join(temp_dir, 'subdir')
os.makedirs(subdir)

ARGS = {
    'task_path': temp_dir,
    'sufix': 'json'
}
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
MixAdapter Ran
ran on a synthesized input
hyintell/citb/continual_learning/AdapterT5.py
pointer only (licence: NONE) · get_code("59cc5cea5f113a6f")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hyintell/citb 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.

Task Ran
ran on a synthesized input · constructed an object; no result was computed
zhaochenyang20/Prompt2Model-Self-Guide/self-guide/utils/tasks.py
code served (permissive licence) · get_code("588da051624b9d01")
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 an instance of its own class, Task: the run constructed the object and did not call it, so no result was computed. 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)

# Define ARGS dict with all required parameters for Task.__init__
ARGS = {
    'task_instruction': 'This is a task instruction for sentiment analysis.',
    'task_name': 'sentiment_classification',
    'examples': [
        {'text': 'This movie is great!', 'label': 'positive'},
        {'text': 'I did not like it.', 'label': 'negative'},
    ],
    'optional_list': ['option1', 'option2', 'option3'],
    'metric': 'exact_match',
    'labels': ['positive', 'negative'],
    'is_classification': True,
}
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 Task — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Task 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
build_dataset_from_file Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/run_ppo.py
code served (permissive licence) · get_code("2d44339dc4e6de45")
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 Dataset. 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 json
import tempfile
import os
from datasets import Dataset

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def __call__(self, text, truncation=True):
        # Return a minimal token structure
        return {'input_ids': [101, 2054, 2003, 102]}

# Create a minimal chat template stand-in
class MinimalChatTemplate:
    def __init__(self):
        self.messages = []
        self.roles = ['user', 'assistant']
    
    def append_message(self, role, content):
        self.messages.append({'role': role, 'content': content})
    
    def get_prompt(self):
        # Return a simple formatted prompt
        return "user: Hello\nassistant: "

# Create temporary JSON dataset file
temp_dir = tempfile.mkdtemp()
dataset_path = os.path.join(temp_dir, 'test_dataset.json')

# Create sample data with required 'input' and 'output' fields
sample_data = [
    {'input': 'What is AI?', 'output': 'AI is artificial intelligence.', 'qtype': 'definition'},
    {'input': 'How does ML work?', 'output': 'ML uses algorithms to learn from data.', 'qtype': 'explanation'},
]

with open(dataset_path, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

# Construct ARGS
ARGS = {
    'tokenizer': MinimalTokenizer(),
    'dataset_path': dataset_path,
    'chat_template': MinimalChatTemplate(),
    'max_sequence_length': 512,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Dataset — a Dataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Dataset 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
collator Ran
ran; our contract draft was wrong, not the code
naver-ai/rethinking-proxy-reward/run_ppo.py
code served (permissive licence) · get_code("bda36ad0c44dbd74")
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) 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)

# The collator function takes a list of dictionaries and groups values by key
# It expects data to be a list where each element is a dict with the same keys

# Create sample data - a list of dicts with consistent keys
data = [
    {
        "input_ids": torch.tensor([1, 2, 3, 4]),
        "attention_mask": torch.tensor([1, 1, 1, 0]),
        "labels": 0
    },
    {
        "input_ids": torch.tensor([5, 6, 7, 8]),
        "attention_mask": torch.tensor([1, 1, 1, 1]),
        "labels": 1
    }
]

ARGS = {
    "data": data
}
Setup block drafted by
a language model, Claude Haiku 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 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
construct_prompt Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/evaluation.py
code served (permissive licence) · get_code("308ab826d7562425")
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

class MessageTemplate:
    def __init__(self):
        self.messages = []
        self.roles = ["user", "assistant"]
    
    def append_message(self, role, message):
        self.messages.append({"role": role, "message": message})
    
    def get_prompt(self):
        prompt_text = ""
        for msg in self.messages:
            if msg["message"] is not None:
                prompt_text += f"{msg['role']}: {msg['message']}\n"
            else:
                prompt_text += f"{msg['role']}: \n"
        return prompt_text

template = MessageTemplate()
msg = "Hello, how are you?"

ARGS = {
    "template": template,
    "msg": msg
}
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
exact_match_score Ran
ran, violated the contract we drafted
renzelou/pick-rank/src/compute_metrics.py
code served (permissive licence) · get_code("e168ce39d04b76f1")
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 string

ARGS = {
    "prediction": "Hello, World!",
    "ground_truth": "hello world",
    "xlingual": False,
}
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 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
get_avg Ran
ran on a synthesized input
renzelou/pick-rank/read_results.py
code served (permissive licence) · get_code("9245d1fb6763d365")
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 float64, ndim 0, 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 numpy as np

# Construct the all_results dictionary with the expected structure
all_results = {
    'base': {
        'RougeL': {'AVG': 0.5},
        'EM': {'AVG': 0.6}
    },
    'model_1': {
        'RougeL': {'AVG': 0.65},
        'EM': {'AVG': 0.70}
    },
    'model_2': {
        'RougeL': {'AVG': 0.72},
        'EM': {'AVG': 0.75}
    },
    'model_3': {
        'RougeL': {'AVG': 0.68},
        'EM': {'AVG': 0.73}
    }
}

ARGS = {
    'all_results': all_results
}
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 float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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
normalize_answer Ran
ran; our contract draft was wrong, not the code
renzelou/pick-rank/src/compute_metrics.py
code served (permissive licence) · get_code("52a9f448d8cbd82a")
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 string

def normalize_answer(s):
    """Lower text and remove punctuation, and extra whitespace."""

    def white_space_fix(text):
        return ' '.join(text.split())

    def remove_punc(text):
        exclude = set(string.punctuation)
        return ''.join((ch for ch in text if ch not in exclude))

    def lower(text):
        return text.lower()
    return white_space_fix(remove_punc(lower(s)))

ARGS = {
    "s": "Hello, World!  This is a TEST."
}
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
read_exp_results Ran
ran on a synthesized input
renzelou/pick-rank/read_results.py
code served (permissive licence) · get_code("60e0493457ee2cde")
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 json
import numpy as np
import tempfile
import os

# Create a temporary JSON file with valid experiment results
temp_dir = tempfile.mkdtemp()
file_path = os.path.join(temp_dir, "results.json")

# Create sample data matching the expected structure
sample_data = {
    "predict_exact_match_for_answerability_classification": 0.85,
    "predict_exact_match_for_cause_effect_classification": 0.78,
    "predict_exact_match_for_coreference_resolution": 0.92,
    "predict_exact_match_for_dialogue_act_recognition": 0.81,
    "predict_exact_match_for_grammar_error_correction": 0.88,
    "predict_exact_match_for_textual_entailment": 0.79,
    "predict_rougeL_for_data_to_text": 0.45,
    "predict_rougeL_for_grammar_error_correction": 0.52,
    "predict_rougeL_for_keyword_tagging": 0.63,
    "predict_rougeL_for_overlap_extraction": 0.71,
    "predict_rougeL_for_question_rewriting": 0.58,
    "predict_rougeL_for_title_generation": 0.67,
}

# Write the sample data to the temporary file
with open(file_path, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

ARGS = {
    "file_name": file_path
}
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 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
read_exp_results_v2 Ran
ran on a synthesized input
renzelou/pick-rank/read_results.py
code served (permissive licence) · get_code("e7f096a075c4af3c")
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 json
import tempfile
import os

# Create a temporary JSON file with the expected structure
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "exp_results.json")

# Prepare the data structure expected by read_exp_results_v2
data = {
    'predict_exact_match_avg': 0.75,
    'predict_rougeL_avg': 0.82,
    'predict_rougeL': [0.80, 0.84, 0.81, 0.83]
}

# Write the JSON file
with open(temp_file, 'w', encoding='utf-8') as f:
    json.dump(data, f)

ARGS = {
    'file_name': temp_file
}
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 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
rouge Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/evaluation.py
code served (permissive licence) · get_code("b00ec969ab1d91fa")
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

# The rouge function takes two string parameters: prediction and ground_truth
# It calls default_rouge_scorer.score() which expects string inputs
# The xlingual parameter is not used in the function body

ARGS = {
    "prediction": "The quick brown fox jumps over the lazy dog",
    "ground_truth": "A quick brown fox leaps over a lazy dog",
    "xlingual": False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, rouge_score, 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
run_superni Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/evaluation.py
code served (permissive licence) · get_code("ee54d5b1181c1329")
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
import numpy as np
from torch.utils.data import TensorDataset, DataLoader
import tempfile
import os
import json

torch.manual_seed(0)

# Create a minimal model that supports generate() and device attribute
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
        self.device_attr = 'cpu'
    
    @property
    def device(self):
        return self.device_attr
    
    def generate(self, input_ids, attention_mask=None, do_sample=False, 
                 eos_token_id=None, pad_token_id=None, max_new_tokens=128):
        # Return dummy output_ids (same shape as input with some extra tokens)
        batch_size = input_ids.shape[0]
        seq_len = input_ids.shape[1]
        # Generate output with max_new_tokens appended
        output = torch.cat([input_ids, torch.ones(batch_size, 10, dtype=torch.long)], dim=1)
        return output

# Create a minimal tokenizer
class MinimalTokenizer:
    def __init__(self):
        self.padding_side = 'left'
        self.eos_token_id = 2
        self.pad_token_id = 0
    
    def __call__(self, texts, return_tensors=None, padding=None, max_length=None, truncation=False):
        # Return dummy token tensors
        batch_size = len(texts)
        seq_len = 12
        input_ids = torch.randint(3, 100, (batch_size, seq_len), dtype=torch.long)
        attention_mask = torch.ones(batch_size, seq_len, dtype=torch.long)
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask
        }
    
    def batch_decode(self, token_ids, skip_special_tokens=False):
        # Return dummy decoded strings
        batch_size = token_ids.shape[0]
        return [f"dummy output {i}" for i in range(batch_size)]

# Create a minimal template
class MinimalTemplate:
    def __init__(self):
        self.roles = ['user', 'assistant']
        self.messages = []
    
    def append_message(self, role, msg):
        self.messages.append({'role': role, 'content': msg})
    
    def get_prompt(self):
        return "dummy prompt"

# Create temporary directory structure
temp_dir = tempfile.mkdtemp()
os.makedirs(f'{temp_dir}/splits/default', exist_ok=True)
os.makedirs(f'{temp_dir}/tasks', exist_ok=True)

# Write test_tasks.txt
with open(f'{temp_dir}/splits/default/test_tasks.txt', 'w') as f:
    f.write('task1\ntask2\n')

# Write dummy task files
task_data = {
    'Definition': ['This is a test task definition.'],
    'Instances': [
        {'input': 'Input 1', 'output': ['Output 1']},
        {'input': 'Input 2', 'output': ['Output 2']},
    ]
}
with open(f'{temp_dir}/tasks/task1.json', 'w') as f:
    json.dump(task_data, f)
with open(f'{temp_dir}/tasks/task2.json', 'w') as f:
    json.dump(task_data, f)

# Construct arguments
ARGS = {
    'model': MinimalModel(),
    'tokenizer': MinimalTokenizer(),
    'TEMPLATE': MinimalTemplate(),
    'target_dir': temp_dir,
    'device': 0,
    'batch_size': 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, rouge_score, 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
set_argument_value Ran
ran; our contract draft was wrong, not the code
yizhongw/tk-instruct/src/create_exps.py
code served (permissive licence) · get_code("d98382bbc755ff32")
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)

# The function set_argument_value operates on a list of arguments
# It looks for a name parameter in the list, and sets the value at the next index
# It checks that the element after the name is not a string starting with '-'
# This appears to be parsing command-line style arguments

# Construct a valid arguments list (like command-line arguments)
arguments = ['--learning_rate', '0.001', '--batch_size', '32', '--model', 'resnet']

# The name to look up
name = '--batch_size'

# The new value to set
value = '64'

ARGS = {
    'arguments': arguments,
    'name': name,
    'value': value
}
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
significant Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/utils.py
code served (permissive licence) · get_code("302783d5f09f24b0")
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 math
import torch
from numbers import Number

torch.manual_seed(0)

# Test with a regular number
x_number = 3.14159

# Test with a tensor
x_tensor = torch.tensor(2.71828)

# Test with zero
x_zero = 0

# Test with NaN
x_nan = float('nan')

# Test with small number
x_small = 0.00456

# Test with large number
x_large = 12345.6789

ARGS = {
    "x": x_number,
    "ndigits": 2
}
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 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
strip_response_tensors Ran
ran on a synthesized input
naver-ai/rethinking-proxy-reward/utils.py
code served (permissive licence) · get_code("f008f637b646cb1f")
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 re

class StandInTokenizer:
    def __init__(self):
        self.padding_side = 'left'
        self.eos_token_id = 2
        self.pad_token_id = 0
    
    def batch_decode(self, tensor, skip_special_tokens=True):
        # Return a list of decoded strings matching batch size
        batch_size = tensor.shape[0]
        return [f"response {i} text" for i in range(batch_size)]
    
    def __call__(self, texts, padding='longest', max_length=768, add_special_tokens=False, return_tensors='pt'):
        # Return tokenized output with input_ids
        batch_size = len(texts)
        # Create varying length sequences
        sequences = []
        for i, text in enumerate(texts):
            seq_len = min(5 + i, max_length)
            seq = torch.randint(1, 10, (seq_len,))
            sequences.append(seq)
        
        # Pad to longest
        max_len = max(len(s) for s in sequences)
        padded = torch.full((batch_size, max_len), self.pad_token_id, dtype=torch.long)
        for i, seq in enumerate(sequences):
            padded[i, :len(seq)] = seq
        
        class TokenizerOutput:
            def __init__(self, input_ids):
                self.input_ids = input_ids
        
        return TokenizerOutput(padded)

torch.manual_seed(0)

# Create response tensors: batch_size=2, sequence_length=4
response_tensors = torch.tensor([
    [1, 5, 3, 2],
    [4, 6, 2, 0]
], dtype=torch.long)

# Create tokenizer
tokenizer = StandInTokenizer()

# Create strip_strings list
strip_strings = ["stop", "end"]

# Create max_length
max_length = 768

ARGS = {
    'response_tensors': response_tensors,
    'tokenizer': tokenizer,
    'strip_strings': strip_strings,
    'max_length': max_length
}
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 — 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
AdapterT5 Not yet run
unverified (harvested, no recorded run)
hyintell/citb/continual_learning/AdapterT5.py
pointer only (licence: NONE) · get_code("7b12231c5a88204c")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hyintell/citb 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. No invocation recorded.

T5AdapterStack Not yet run
unverified (harvested, no recorded run)
hyintell/citb/continual_learning/AdapterT5.py
pointer only (licence: NONE) · get_code("b95b7688a76ec173")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, hyintell/citb 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. No invocation recorded.

create_task Not yet run
unverified (harvested, no recorded run)
allenai/natural-instructions-expansion/src/convert_prompt.py
code served (permissive licence) · get_code("d124ae4da8ffcd68")
How it ran

no invocation recorded.

del_stop_words Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/edition_slight.py
code served (permissive licence) · get_code("8ee347f40e852033")
How it ran

no invocation recorded.

del_words Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/edition_slight.py
code served (permissive licence) · get_code("151f29fd0f845085")
How it ran

no invocation recorded.

get_conv_template Not yet run
unverified (harvested, no recorded run)
naver-ai/rethinking-proxy-reward/conversation.py
code served (permissive licence) · get_code("a3ac8409c5a9506e")
How it ran

no invocation recorded.

load_tf_weights_in_t5 Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/src/modeling_pointer.py
code served (permissive licence) · get_code("edfe67269d5fa277")
How it ran

no invocation recorded.

randomly_remain_sen Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/edition_huge.py
code served (permissive licence) · get_code("616bd4384ae83258")
How it ran

no invocation recorded.

remain_first_sen Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/edition_huge.py
code served (permissive licence) · get_code("89a91b4883ca9a87")
How it ran

no invocation recorded.

save_json Not yet run
unverified (harvested, no recorded run)
allenai/natural-instructions-expansion/src/convert_prompt.py
code served (permissive licence) · get_code("dd7ca8d7f3539848")
How it ran

no invocation recorded.

split_sen Not yet run
unverified (harvested, no recorded run)
renzelou/pick-rank/edition_huge.py
code served (permissive licence) · get_code("182ec5854420893c")
How it ran

no invocation recorded.

Also present in other papers' code

Of 28 samples lifted from this paper's linked repositories, 28 carry a function identity and 0 do not. Of the 28 distinct functions with an identity, 6 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 6 later) and 22 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
get_conv_template 69 other papers
2 dated earlier, 66 later, 1 undated
1911.03705 (2019-11)
CommonGen: A Constrained Text Generation Challenge for Generative Commonsense Reasoning
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2302.13971 (2023-02)
LLaMA: Open and Efficient Foundation Language Models
and 66 more not named here
load_tf_weights_in_t5 20 other papers
1 dated earlier, 19 later
2201.05966 (2022-01)
UNIFIEDSKG: Unifying and Multi-Tasking Structured Knowledge Grounding with Text-to-Text Language Models
2205.06983 (2022-05)
RASAT: Integrating Relational Structures into Pretrained Seq2Seq Model for Text-to-SQL
2209.14627 (2022-09)
An Equal-Size Hard EM Algorithm for Diverse Dialogue Generation
and 17 more not named here
normalize_answer 12 other papers
12 later
2212.03813 (2022-12)
Robustness of Learning from Task Instructions
2023.findings-emnlp.836 (2023 (year only))
Ensemble-Instruct: Generating Instruction-Tuning Data with a Heterogeneous Mixture of LMs
2302.14691 (2023-02)
Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following
and 9 more not named here
collator 9 other papers
1 dated earlier, 8 later
1909.08593 (2019-09)
Fine-Tuning Language Models from Human Preferences
2304.08177 (2023-04)
Efficient and Effective Text Encoding for Chinese LLaMA and Alpaca
2309.16155 (2023-09)
The Trickle-down Impact of Reward (In-)consistency on RLHF
and 6 more not named here
exact_match_score 8 other papers
8 later
2212.03813 (2022-12)
Robustness of Learning from Task Instructions
2302.14691 (2023-02)
Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following
2305.00944 (2023-05)
Poisoning Language Models During Instruction Tuning
and 5 more not named here
set_argument_value 2 other papers
2 later
2212.03813 (2022-12)
Robustness of Learning from Task Instructions
2406.12382 (2024-06)
From Instance Training to Instruction Learning: Task Adapters Generation from Instructions

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

How well can NLP models generalize to a variety of unseen tasks when provided with task instructions? To address this question, we first introduce SUPER-NATURALINSTRUCTIONS, 1 a benchmark of 1,616 diverse NLP tasks and their expert-written instructions. Our collection covers 76 distinct task types, including but not limited to classification, extraction, infilling, sequence tagging, text rewriting, and text composition. This large and diverse collection of tasks enables rigorous benchmarking of cross-task generalization under instructionstraining models to follow instructions on a subset of tasks and evaluating them on the remaining unseen ones. Furthermore, we build Tk-INSTRUCT, a transformer model trained to follow a variety of in-context instructions (plain language task definitions or k-shot examples). Our experiments show that Tk-INSTRUCT outperforms existing instruction-following models such as Instruct-GPT by over 9% on our benchmark despite being an order of magnitude smaller. We further analyze generalization as a function of various scaling parameters, such as the number of observed tasks, the number of instances per task, and model sizes. We hope our dataset and model facilitate future progress towards more general-purpose NLP models. 2

For agents

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

get_harvested_code_for_paper("2204.07705")
get_code_for_paper("2204.07705")
have("2204.07705")

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

curl https://syntology.ai/api/ran/2204.07705.json

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

[![Syntology run record](https://syntology.ai/api/ran/2204.07705.svg)](https://syntology.ai/paper/2204.07705)

Connect an agent — have() is free.