SciFactCheck

SciFactCheck is a benchmark and evaluation framework for assessing scientific factuality hallucinations in open-ended, closed-book LLM-generated text across five scientific domains: Engineering, Life Sciences, Physical Sciences and Mathematics, Social and Behavioural Sciences, and Arts and Humanities. The framework targets three hallucination types: unverifiability (VER), overclaim (OVR), and attribution (ATR), with dedicated automatic metrics per type.

This repository contains the code for benchmark construction, model prompting, and evaluation.

Repository Structure

SciFactCheck/
├── data_prep/          # Benchmark construction + utils
├── prompting/          # Model inference
└── evaluation/         # Hallucination evaluation

1. Benchmark Construction

The benchmark consists of 2,500 scientific concepts across five domains, grounded in peer-reviewed survey literature. Concepts are extracted semi-automatically from highly cited (>=100 citations) survey papers retrieved from S2ORC, and from the Internet Encyclopedia of Philosophy (IEP) for Arts and Humanities.

• S2ORC survey paper retrieval: data_prep/s2orc_api.py
• IEP article scraping: data_prep/iep.py
• Full dataset: scifactcheck_data.csv, also available on HuggingFace: https://huggingface.co/datasets/rabuahmad/scifactcheck

2. Model Prompting

We use a two-step prompting strategy: first generating a scientific paragraph, then eliciting citations separately.

• Paragraph generation prompt: prompting/prompt_paragraph.txt
• Citation generation prompt: prompting/prompt_citation.txt
• Inference script: prompting/run_model.py

The inference script supports three backends: HuggingFace (hf), LiteLLM (litellm), and vLLM (vllm). API keys can be passed as arguments or set as environment variables (HF_API_KEY, LITELLM_API_KEY).

Basic usage (single prompt)

python prompting/run_model.py \
    --model_name Qwen/Qwen2.5-7B-Instruct \
    --model_type hf \
    --dataset_name rabuahmad/scifactcheck \
    --prompt_template prompting/prompt_paragraph.txt \
    --output_dir results/ \
    --max_new_tokens 512 \
    --apply_chat_template

Two-prompt strategy (paragraph + citations)

python prompting/run_model.py \
    --model_name Qwen/Qwen2.5-7B-Instruct \
    --model_type hf \
    --dataset_name rabuahmad/scifactcheck \
    --prompt_template prompting/prompt_paragraph.txt \
    --two_prompts \
    --citation_prompt_template prompting/prompt_citation.txt \
    --output_dir results/ \
    --apply_chat_template

Resuming from a checkpoint

python prompting/run_model.py \
    --model_name Qwen/Qwen2.5-7B-Instruct \
    --model_type vllm \
    --tensor_parallel_size 4 \
    --dataset_name rabuahmad/scifactcheck \
    --prompt_template prompting/prompt_paragraph.txt \
    --two_prompts \
    --citation_prompt_template prompting/prompt_citation.txt \
    --from_checkpoint \
    --checkpoint_path results/Qwen_Qwen2.5-7B-Instruct_checkpoint.jsonl \
    --output_dir results/

All arguments

Argument	Required	Default	Description
--model_name	✓	—	Model name or HuggingFace path
--model_type	✓	—	Backend: hf, litellm, or vllm
--dataset_name	✓	—	HuggingFace dataset identifier (SciFactCheck)
--prompt_template	✓	—	Path to .txt prompt template; use {scientific_entity} and {field} as placeholders
--output_dir		results/	Directory to save outputs
--max_new_tokens		512	Maximum number of tokens to generate
--tensor_parallel_size		1	Number of tensor parallel partitions (vLLM only)
--apply_chat_template		False	Format prompts with tokenizer.apply_chat_template()
--two_prompts		False	Use two-prompt strategy (paragraph + citations)
--citation_prompt_template		—	Path to citation prompt template (required if --two_prompts)
--from_checkpoint		False	Resume from a previous checkpoint
--checkpoint_path		—	Path to checkpoint file (required if --from_checkpoint)
--limit		None	Limit number of examples (for testing)
--hf_api_key		$HF_API_KEY	HuggingFace API key
--litellm_api_key		$LITELLM_API_KEY	LiteLLM API key

Sample output format: one record per concept per model would look like this:

{
    "id": 0,
    "field": "Engineering"
    "entity": "nanofluids",
    "two_prompts": true,
    "output": {
        "prompt": "Scientific claims are those that [...]",
        "text": "Nanofluids are engineered colloidal suspensions that [...]",
        "logprobs": [...],
        "length_norm_logprob": -0.6522063442125952,
    }
    "meta": {
        "model": "openai/gpt-4o-mini",
        "backend": "litellm",
        "num_tokens": 249,
    }
    "output_citation": {
        "prompt": "Find relevant scientific or academic [...]",
        "text": "Hassan, H. A., & Badran, M. A. (2016). [...]",
        "logprobs": [...],
        "length_norm_logprob": -219.88604030781042,
        "meta": {
            "model": "openai/gpt-4o-mini",
            "backend": "litellm",
            "num_tokens": 319
        }
    }
}

3. Evaluation Framework

3.1 Unverifiability Hallucination (VER)

We use two metrics with different knowledge sources:

OpenFActScore (OFS) verifies claims against the full text of the source survey paper.

• KB construction code: data_prep/prepare_ofs_kb.py
• We use a forked version of OpenFActScore: https://github.com/ryabhmd/OpenFActScore

VeriScore verifies claims against Google Search results via the Serper API.

• We use a forked version of VeriScore: https://github.com/ryabhmd/VeriScore

Both metrics use VeriScore's claim extraction model for consistency.

3.2 Overclaim Hallucination (OVR)

Two complementary uncertainty proxies are computed:

• Linguistic uncertainty (certainty estimator): sentence-level certainty scores
• Parametric uncertainty (log probabilities): length-normalized mean per-token log probability

3.3 Attribution Hallucination (ATR)

Citations are verified against CrossRef using DOI-based and title-based lookup. A citation is deemed valid if either strategy confirms it.

• CrossRef validator: evaluation/citation_validator.py

OVR and ATR metrics can be run directly using evaluation/run_eval.py.

Basic usage

python evaluation/run_eval.py \
    --model_name Qwen/Qwen2.5-7B-Instruct \
    --file results/Qwen_Qwen2.5-7B-Instruct.jsonl \
    --save results/eval/ \
    --eval ovr-uncertainty ovr-logprobs attr

With citation parsing

If your model output contains raw citation strings that need to be parsed before ATR evaluation, add --parse_citations:

python evaluation/run_eval.py \
    --model_name Qwen/Qwen2.5-7B-Instruct \
    --file results/Qwen_Qwen2.5-7B-Instruct.jsonl \
    --save results/eval/ \
    --eval atr \
    --parse_citations

All arguments

Argument	Required	Default	Description
--model_name	✓	—	Model name or HuggingFace path
--file	✓	—	Path to model output JSONL file
--save	✓	—	Directory to save evaluation results
--eval	✓	—	One or more metrics to evaluate: ovr-uncertainty (linguistic uncertainty), ovr-logprobs (parametric uncertainty), attr (attribution validation)
--parse_citations		False	Parse raw citation strings before ATR evaluation

Expected output format: one record per concept:

{
    "id": 0,
    "field": "Engineering",
    "entity": "nanofluids",
    "two_prompts": true,
    "output": {
        "prompt": "...",
        "text": "...",
        "logprobs": [...],
        "length_norm_logprob": -0.652,
        "meta": {
            "model": "openai/gpt-4o-mini",
            "backend": "litellm",
            "num_tokens": 249
        }
    },
    "output_citation": {
        "prompt": "...",
        "text": "...",
        "logprobs": [...],
        "length_norm_logprob": -219.88,
        "meta": {
            "model": "openai/gpt-4o-mini",
            "backend": "litellm",
            "num_tokens": 319
        },
        "parsed_citations": [
            {
                "authors": "Choi, S. U. S., & Eastman, J. A",
                "year": "1995",
                ...
            }
        ]
    },
    "eval_logprobs": -0.652,
    "eval_uncertainty": {
        "mean_sentence_certainty": 4.77,
        "aspect_scores": {
            "Probability": 0.67,
            "Number": 0.0,
            "Extent": 0.0,
            "Condition": null,
            "Framing": 1.0,
            "Suggestion": 1.0
        }
    }
}