SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2203.12990 · ACL · 2022

Generating Scientific Claims for Zero-Shot Scientific Fact Checking

Arman Cohan, Kyle Lo, David Wadden, Dustin Wright, Isabelle Augenstein, Lucy Wang, Bailey Kuehl

arXiv · PDF · Open in the Atlas

Code that ran

6 of 7 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; 0 honoured the contract we drafted, 0 violated it and 6 have no contract verdict; 0 failures of our instrument; 1 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/2203.12990.json · badge: /api/ran/2203.12990.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
allenai/scientific-claim-generation canonical (classifier label)
classifier basis not recorded on the link
6 ran · 1 no recorded run
FunctionStatusWhere it lives
compute_metrics Ran
ran on a synthesized input
allenai/scientific-claim-generation/archive/run_bart_claim_generation.py
code served (permissive licence) · get_code("e24368339d19b559")
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 numpy as np
from unittest.mock import Mock

# Create a minimal tokenizer that supports batch_decode
class MinimalTokenizer:
    def __init__(self):
        self.pad_token_id = 0
    
    def batch_decode(self, token_ids, skip_special_tokens=True):
        # Convert token ids to strings
        decoded = []
        for ids in token_ids:
            decoded.append(' '.join(str(int(id_)) for id_ in ids if int(id_) != self.pad_token_id))
        return decoded

# Create a minimal metric object that returns rouge scores
class MinimalRougeScore:
    def __init__(self):
        self.low = Mock(precision=0.85, recall=0.80, fmeasure=0.825)
        self.mid = Mock(precision=0.87, recall=0.82, fmeasure=0.845)
        self.high = Mock(precision=0.90, recall=0.85, fmeasure=0.875)

class MinimalMetric:
    def compute(self, predictions, references):
        return {
            'rouge1': MinimalRougeScore(),
            'rouge2': MinimalRougeScore(),
            'rougeL': MinimalRougeScore(),
            'rougeLsum': MinimalRougeScore(),
        }

# Create sample predictions and labels
# Batch size 2, sequence length 3
preds = np.array([
    [1, 2, 3],
    [4, 5, 6]
], dtype=np.int64)

labels = np.array([
    [1, 2, -100],
    [4, -100, 6]
], dtype=np.int64)

eval_preds = (preds, labels)

ARGS = {
    'tokenizer': MinimalTokenizer(),
    'metric': MinimalMetric(),
    'eval_preds': eval_preds,
}
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
data_preprocess Ran
ran on a synthesized input
allenai/scientific-claim-generation/archive/run_bart_claim_generation.py
code served (permissive licence) · get_code("f6a48a7da22d1cc1")
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
from transformers import AutoTokenizer

torch.manual_seed(0)

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def __init__(self):
        self.model_max_length = 128
        self._target_mode = False
    
    def __call__(self, texts, max_length=None, truncation=False):
        # Return minimal tokenizer output structure
        input_ids = [[101] * min(len(text.split()), max_length or 128) for text in texts]
        return {
            'input_ids': input_ids,
            'attention_mask': [[1] * len(ids) for ids in input_ids]
        }
    
    def as_target_tokenizer(self):
        return self
    
    def __enter__(self):
        self._target_mode = True
        return self
    
    def __exit__(self, *args):
        self._target_mode = False

tokenizer = MinimalTokenizer()

# Create examples for 'citeworth' dataset
examples_citeworth = {
    'generated_question': ['What is AI?', 'What is ML?'],
    'answer': [
        {'text': 'Artificial Intelligence'},
        {'text': 'Machine Learning'}
    ]
}

# Create examples for non-citeworth dataset
examples_other = {
    'question': ['What is AI?', 'What is ML?'],
    'answer': ['Artificial Intelligence', 'Machine Learning'],
    'turker_answer': ['AI is a field', 'ML is a subset']
}

ARGS = {
    'tokenizer': tokenizer,
    'dset': 'citeworth',
    'examples': examples_citeworth
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — 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
data_preprocess Ran
ran on a synthesized input
allenai/scientific-claim-generation/archive/run_bart_scifact_claim_generation.py
code served (permissive licence) · get_code("174036c49c9e752a")
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
from unittest.mock import Mock

torch.manual_seed(0)

class StandInTokenizer:
    def __init__(self):
        self.model_max_length = 128
    
    def __call__(self, texts, max_length=None, truncation=False):
        # Return a dict with input_ids
        # Simulate tokenization by creating dummy token sequences
        result = {
            'input_ids': [[101, 2054, 2003] + [0] * (max_length - 3) if max_length else [101, 2054, 2003] for _ in texts],
            'attention_mask': [[1] * 3 + [0] * (max_length - 3) if max_length else [1, 1, 1] for _ in texts]
        }
        return result
    
    def as_target_tokenizer(self):
        # Return a context manager that does nothing
        return self

    def __enter__(self):
        return self
    
    def __exit__(self, *args):
        pass

tokenizer = StandInTokenizer()

examples = {
    'context': [
        'The sky is blue.',
        'Water flows downhill.'
    ],
    'claims': [
        'The sky is blue.',
        'Water flows downhill.'
    ]
}

dset = None

ARGS = {
    'tokenizer': tokenizer,
    'dset': dset,
    'examples': examples
}
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
q2c_data_preprocess Ran
ran on a synthesized input
allenai/scientific-claim-generation/claimgen_entity.py
code served (permissive licence) · get_code("5b3466b7e8552a30")
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 numpy as np
from collections import defaultdict

class MockTokenizer:
    def __init__(self):
        self.model_max_length = 512
    
    def __call__(self, texts, max_length=512, truncation=True):
        # Return a dict with input_ids and attention_mask
        # Simple mock: just create token ids based on text length
        input_ids = []
        for text in texts:
            # Simple tokenization: split by spaces and create token ids
            tokens = text.split() if text else []
            ids = [i % 100 + 1 for i in range(len(tokens))][:max_length]
            # Pad to same length
            ids = ids + [0] * (max_length - len(ids))
            input_ids.append(ids)
        
        return {
            'input_ids': input_ids,
            'attention_mask': [[1 if x != 0 else 0 for x in ids] for ids in input_ids]
        }
    
    def as_target_tokenizer(self):
        return self

    def __enter__(self):
        return self
    
    def __exit__(self, *args):
        pass

# Test with citeworth dataset
tokenizer_citeworth = MockTokenizer()

examples_citeworth = {
    'generated_question': ['What is the capital?', 'Who wrote this?'],
    'answer': [
        {'text': 'Paris'},
        {'text': 'Shakespeare'}
    ]
}

# Test with squad-like dataset
tokenizer_squad = MockTokenizer()

examples_squad = {
    'question': ['What is the capital of France?', 'Who wrote Hamlet?'],
    'answer': ['Paris', 'Shakespeare'],
    'turker_answer': ['The capital is Paris', 'Shakespeare wrote it']
}

ARGS = {
    'tokenizer': tokenizer_squad,
    'dset': 'squad',
    'examples': examples_squad
}
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 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
qg_data_preprocess Ran
ran on a synthesized input
allenai/scientific-claim-generation/claimgen_entity.py
code served (permissive licence) · get_code("eebb1231ae66ae99")
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
from transformers import AutoTokenizer

torch.manual_seed(0)

# Create a minimal tokenizer mock that behaves like a HuggingFace tokenizer
class MockTokenizer:
    def __init__(self):
        self.model_max_length = 512
        self._is_target = False
    
    def __call__(self, texts, max_length=None, truncation=False):
        # Return a dict with input_ids and attention_mask
        if isinstance(texts, str):
            texts = [texts]
        batch_size = len(texts)
        return {
            'input_ids': [[101] * 10 for _ in range(batch_size)],
            'attention_mask': [[1] * 10 for _ in range(batch_size)]
        }
    
    def as_target_tokenizer(self):
        return self
    
    def __enter__(self):
        self._is_target = True
        return self
    
    def __exit__(self, *args):
        self._is_target = False

tokenizer = MockTokenizer()

# Create example data for 'local' dataset
examples_local = {
    'context': ['The cat sat on the mat', 'Dogs are loyal animals'],
    'answers': [
        [{'text': 'cat'}],
        [{'text': 'dogs'}]
    ],
    'question': ['What animal sat on the mat?', 'What animals are loyal?']
}

# Create example data for non-local dataset (e.g., squad format)
examples_squad = {
    'context': ['Paris is the capital of France', 'The Eiffel Tower is in Paris'],
    'answers': [
        {'text': ['France']},
        {'text': ['Paris']}
    ],
    'question': ['What is the capital of France?', 'Where is the Eiffel Tower?']
}

ARGS = {
    'tokenizer': tokenizer,
    'dset': 'local',
    'examples': examples_local
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — 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
sort_fc_claims Ran
ran on a synthesized input
allenai/scientific-claim-generation/claimgen_bart.py
code served (permissive licence) · get_code("6c83dbde7a7359ce")
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 scipy.special import softmax

torch.manual_seed(0)

# Construct preds: list of dicts with 'id' and 'evidence' keys
# Each evidence value has a 'score' key with logits
preds = [
    {
        'id': 'claim_0',
        'evidence': {
            'evidence_0': {'score': np.array([0.1, 2.0, 1.5])},
            'evidence_1': {'score': np.array([0.5, 1.2, 0.8])}
        }
    },
    {
        'id': 'claim_1',
        'evidence': {
            'evidence_2': {'score': np.array([1.0, 0.5, 1.8])},
            'evidence_3': {'score': np.array([0.2, 0.9, 0.3])}
        }
    }
]

# Construct original_claims: list of dicts with 'id' key
original_claims = [
    {'id': 'claim_0', 'text': 'First claim'},
    {'id': 'claim_1', 'text': 'Second claim'}
]

ARGS = {
    'preds': preds,
    'original_claims': original_claims
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, 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_perplexity Not yet run
unverified (harvested, no recorded run)
allenai/scientific-claim-generation/generate_claim_variants.py
code served (permissive licence) · get_code("75bf33d25ca9a6cd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 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 7 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

Automated scientific fact checking is difficult due to the complexity of scientific language and a lack of significant amounts of training data, as annotation requires domain expertise. To address this challenge, we propose scientific claim generation, the task of generating one or more atomic and verifiable claims from scientific sentences, and demonstrate its usefulness in zero-shot fact checking for biomedical claims. We propose CLAIMGEN-BART, a new supervised method for generating claims supported by the literature, as well as KBIN, a novel method for generating claim negations. Additionally, we adapt an existing unsupervised entity-centric method of claim generation to biomedical claims, which we call CLAIMGEN-ENTITY. Experiments on zero-shot fact checking demonstrate that both CLAIMGEN-ENTITY and CLAIMGEN-BART, coupled with KBIN, achieve up to 90% performance of fully supervised models trained on manually annotated claims and evidence. A rigorous evaluation study demonstrates significant improvement in generated claim and negation quality over existing baselines. 1

For agents

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

get_harvested_code_for_paper("2203.12990")
get_code_for_paper("2203.12990")
have("2203.12990")

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

curl https://syntology.ai/api/ran/2203.12990.json

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

[![Syntology run record](https://syntology.ai/api/ran/2203.12990.svg)](https://syntology.ai/paper/2203.12990)

Connect an agent — have() is free.