SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 1903.12136 · 2019

Distilling Task-Specific Knowledge from BERT into Simple Neural Networks

arXiv · PDF · Open in the Atlas

Code that ran

9 of 9 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; 3 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 0 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/1903.12136.json · badge: /api/ran/1903.12136.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
castorini/d-bert reimplementation (classifier label)
classifier basis: classifier_2025
8 ran · 0 no recorded run
pvgladkov/knowledge-distillation reimplementation (classifier label)
classifier basis: classifier_2025
1 ran · 0 no recorded run
FunctionStatusWhere it lives
batch_to_inputs Ran
ran; our contract draft was wrong, not the code
pvgladkov/knowledge-distillation/knowledge_distillation/bert_trainer.py
code served (permissive licence) · get_code("3ea37bfa7bd6625b")
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 function expects batch to be indexable with [0], [1], [2], [3]
# Each element should be a tensor based on the keys being used
# (input_ids, attention_mask, token_type_ids, labels)

batch = [
    torch.randint(0, 100, (2, 4)),  # input_ids: batch_size=2, seq_len=4
    torch.ones((2, 4), dtype=torch.long),  # attention_mask: batch_size=2, seq_len=4
    torch.zeros((2, 4), dtype=torch.long),  # token_type_ids: batch_size=2, seq_len=4
    torch.randint(0, 2, (2,)),  # labels: batch_size=2
]

ARGS = {
    'batch': batch
}
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
convert_dp_to_single Ran
ran; our contract draft was wrong, not the code
castorini/d-bert/dbert/finetune/classifier.py
code served (permissive licence) · get_code("31cb7e5ee37e07ad")
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
import torch.nn as nn

torch.manual_seed(0)

# The function convert_dp_to_single takes a state_dict (dictionary)
# and removes 'module.' prefix from keys that have it.
# This is typically used to convert DataParallel model state dicts to single GPU format.

# Create a sample state_dict with some keys prefixed with 'module.'
state_dict = {
    'module.layer1.weight': torch.randn(4, 3),
    'module.layer1.bias': torch.randn(4),
    'module.layer2.weight': torch.randn(5, 4),
    'layer3.weight': torch.randn(3, 5),  # This one doesn't have 'module.' prefix
    'layer3.bias': torch.randn(3),
}

ARGS = {
    'state_dict': state_dict
}
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
convert_single_to_dp Ran
ran; our contract draft was wrong, not the code
castorini/d-bert/dbert/finetune/classifier.py
code served (permissive licence) · get_code("76d68a1ed9ae4351")
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
import torch.nn as nn

torch.manual_seed(0)

# The function convert_single_to_dp takes a state_dict (dictionary of model parameters)
# and adds 'module.' prefix to all keys that don't already have it.
# This is used to convert a single-GPU model state_dict to DataParallel format.

# Create a simple state_dict with some model parameters
state_dict = {
    'layer1.weight': torch.randn(4, 3),
    'layer1.bias': torch.randn(4),
    'layer2.weight': torch.randn(5, 4),
    'layer2.bias': torch.randn(5),
}

ARGS = {
    'state_dict': state_dict
}
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
evaluate Ran
ran, honoured the contract we drafted
castorini/d-bert/dbert/distill/run/distill_birnn.py
code served (permissive licence) · get_code("502edb147c17e8f6")
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 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 torch.nn as nn

torch.manual_seed(0)

# Build a minimal model that satisfies the function's usage
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)  # 4 input dims, 3 output classes
    
    def forward(self, x):
        return self.linear(x)

# Build a minimal batch class with required attributes
class Batch:
    def __init__(self, sentence, label):
        self.sentence = sentence
        self.label = label

# Build a minimal dataset iterator
class DatasetIterator:
    def __init__(self, batches):
        self.batches = batches
        self.index = 0
    
    def init_epoch(self):
        self.index = 0
    
    def __iter__(self):
        self.index = 0
        return self
    
    def __next__(self):
        if self.index >= len(self.batches):
            raise StopIteration
        batch = self.batches[self.index]
        self.index += 1
        return batch

# Create model
model = MinimalModel()

# Create batches with distinct dimensions
# sentence: (batch_size=2, feature_dim=4)
# label: (batch_size=2,) with class indices in range [0, 2]
batch1 = Batch(
    sentence=torch.randn(2, 4),
    label=torch.tensor([0, 1], dtype=torch.long)
)
batch2 = Batch(
    sentence=torch.randn(2, 4),
    label=torch.tensor([2, 0], dtype=torch.long)
)

# Create dataset iterator
ds_iter = DatasetIterator([batch1, batch2])

# Create criterion
criterion = nn.CrossEntropyLoss()

ARGS = {
    "model": model,
    "ds_iter": ds_iter,
    "criterion": criterion,
    "export_eval_labels": False
}
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 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
evaluate_score Ran
ran, honoured the contract we drafted
castorini/d-bert/dbert/distill/run/distill_birnn.py
code served (permissive licence) · get_code("e4a51df9636b003c")
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 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 torch.nn as nn
from torch.utils.data import DataLoader

torch.manual_seed(0)

# Create a minimal model that takes batch.sentence and returns scores
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 1)
    
    def forward(self, sentence):
        # sentence is expected to be a tensor
        return self.linear(sentence)
    
    def eval(self):
        super().eval()
        return self

# Create a minimal batch class with required attributes
class MinimalBatch:
    def __init__(self, sentence, score):
        self.sentence = sentence
        self.score = score

# Create a minimal dataset iterator with init_epoch method
class MinimalDatasetIterator:
    def __init__(self, batches):
        self.batches = batches
        self.index = 0
    
    def init_epoch(self):
        self.index = 0
    
    def __iter__(self):
        self.index = 0
        return self
    
    def __next__(self):
        if self.index >= len(self.batches):
            raise StopIteration
        batch = self.batches[self.index]
        self.index += 1
        return batch

# Create minimal batches
batch1 = MinimalBatch(
    sentence=torch.randn(2, 4),  # batch_size=2, feature_dim=4
    score=torch.tensor([1.5, 2.5])
)
batch2 = MinimalBatch(
    sentence=torch.randn(2, 4),
    score=torch.tensor([3.5, 4.5])
)

# Create dataset iterator
ds_iter = MinimalDatasetIterator([batch1, batch2])

# Create model
model = MinimalModel()

# Create criterion (not used in function but required parameter)
criterion = nn.MSELoss()

ARGS = {
    "model": model,
    "ds_iter": ds_iter,
    "criterion": criterion,
    "export_eval_labels": False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
scipy, torch, tqdm — 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
init_sos Ran
ran, honoured the contract we drafted
castorini/d-bert/dbert/generate/finetune_transfoxl.py
code served (permissive licence) · get_code("59e0610210960189")
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 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 MockTransformer(nn.Module):
    def __init__(self):
        super().__init__()
        # wte is the word token embedding layer
        # Initialize with a small embedding: vocab_size=10, embedding_dim=4
        self.wte = nn.Embedding(10, 4)

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.transformer = MockTransformer()

model = MockModel()

ARGS = {
    "model": model
}
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 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
mse_forward_monkey_patch Ran
ran; our contract draft was wrong, not the code
castorini/d-bert/dbert/finetune/classifier.py
code served (permissive licence) · get_code("fb341ab6d38412a8")
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 Tensor, 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 torch.nn as nn

torch.manual_seed(0)

# Build a minimal BERT-like model that has the expected interface
class MinimalBERT(nn.Module):
    def __init__(self, hidden_size=4, vocab_size=100):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_size)
        self.encoder = nn.Linear(hidden_size, hidden_size)
        self.pooler = nn.Linear(hidden_size, hidden_size)
    
    def forward(self, input_ids, token_type_ids=None, attention_mask=None, output_all_encoded_layers=False):
        # Returns (sequence_output, pooled_output)
        embedded = self.embedding(input_ids)
        encoded = self.encoder(embedded)
        # Pool by taking mean over sequence dimension
        pooled = self.pooler(encoded.mean(dim=1))
        return encoded, pooled

# Build the model that uses the monkey-patched forward method
class BertForSequenceClassification(nn.Module):
    def __init__(self, num_labels=3, hidden_size=4):
        super().__init__()
        self.num_labels = num_labels
        self.bert = MinimalBERT(hidden_size=hidden_size)
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(hidden_size, num_labels)
    
    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None, mse=False):
        # This will be replaced by the monkey patch, but we define it for completeness
        pass

# Create model instance
model = BertForSequenceClassification(num_labels=3, hidden_size=4)

# Create input tensors
batch_size = 2
seq_length = 5
input_ids = torch.randint(0, 100, (batch_size, seq_length))
token_type_ids = torch.zeros((batch_size, seq_length), dtype=torch.long)
attention_mask = torch.ones((batch_size, seq_length), dtype=torch.long)

# Create labels for MSE loss (continuous values)
labels_mse = torch.randn(batch_size, 1)

# Create labels for CrossEntropy loss (class indices)
labels_ce = torch.randint(0, 3, (batch_size,))

# Build ARGS dict with MSE=True case
ARGS = {
    'self': model,
    'input_ids': input_ids,
    'token_type_ids': token_type_ids,
    'attention_mask': attention_mask,
    'labels': labels_mse,
    'mse': True
}
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 Tensor, 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
sample_query Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
castorini/d-bert/dbert/generate/finetune_transfoxl.py
code served (permissive licence) · get_code("36da55e25256092f")
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 str. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.distributions.categorical import Categorical

torch.manual_seed(0)

class SimpleTokenizer:
    def __init__(self):
        self.vocab = {
            'hello': 0,
            'world': 1,
            'test': 2,
            '<eos>': 3,
            'a': 4,
            'b': 5,
            'c': 6,
            'd': 7,
        }
        self.idx_to_sym = {v: k for k, v in self.vocab.items()}
    
    def tokenize(self, text):
        return text.split()
    
    def get_idx(self, token):
        return self.vocab.get(token, 0)
    
    def get_sym(self, idx):
        return self.idx_to_sym.get(idx, '<unk>')

class SimpleModel(nn.Module):
    def __init__(self, vocab_size=8, hidden_dim=4):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_dim)
        self.linear = nn.Linear(hidden_dim, vocab_size)
    
    def forward(self, input_ids, mems=None):
        x = self.embedding(input_ids)
        x = x.mean(dim=1, keepdim=True)
        logits = self.linear(x)
        present = None
        return logits, present

model = SimpleModel(vocab_size=8, hidden_dim=4)
tokenizer = SimpleTokenizer()
text = "hello world"

ARGS = {
    'model': model,
    'tokenizer': tokenizer,
    'text': text,
    'n': 128,
    'paired': False,
}
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 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
transfo_encode Ran
ran; our contract draft was wrong, not the code
castorini/d-bert/dbert/generate/finetune_transfoxl.py
code served (permissive licence) · get_code("4837663eda94ee99")
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 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

class MockTokenizer:
    def tokenize(self, text, add_eos=False):
        # Simple tokenization: split by spaces and add eos token if requested
        tokens = text.split()
        if add_eos:
            tokens.append('<eos>')
        return tokens
    
    def convert_tokens_to_ids(self, tokens):
        # Simple mapping: each unique token gets an id
        token_to_id = {
            '<eos>': 1,
            'hello': 2,
            'world': 3,
            'test': 4,
            'query': 5,
        }
        return [token_to_id.get(t, 0) for t in tokens]

tokenizer = MockTokenizer()

queries = [
    "hello world",
    "test query here"
]

ARGS = {
    'tokenizer': tokenizer,
    'queries': queries,
    'sos_idx': None,
    'eos': True,
    'max_len': 80,
    'split_encode': False,
    'idx2remap': None,
    'return_raw': False,
    'condition_model': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 9 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.

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

In the natural language processing literature, neural networks are becoming increasingly deeper and complex. The recent poster child of this trend is the deep language representation model, which includes BERT, ELMo, and GPT. These developments have led to the conviction that previous-generation, shallower neural networks for language understanding are obsolete. In this paper, however, we demonstrate that rudimentary, lightweight neural networks can still be made competitive without architecture changes, external training data, or additional input features. We propose to distill knowledge from BERT, a state-of-the-art language representation model, into a single-layer BiLSTM, as well as its siamese counterpart for sentence-pair tasks. Across multiple datasets in paraphrasing, natural language inference, and sentiment classification, we achieve comparable results with ELMo, while using roughly 100 times fewer parameters and 15 times less inference time.

For agents

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

get_harvested_code_for_paper("1903.12136")
get_code_for_paper("1903.12136")
have("1903.12136")

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

curl https://syntology.ai/api/ran/1903.12136.json

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

[![Syntology run record](https://syntology.ai/api/ran/1903.12136.svg)](https://syntology.ai/paper/1903.12136)

Connect an agent — have() is free.