README.md

---
language:
- en
pretty_name: SciFact Chat-Format
license: cc-by-nc-2.0
task_categories:
- text-classification
task_ids:
- fact-checking
source_datasets:
- allenai/scifact
tags:
- text
- science
- claim-verification
- chat-format
- instruction-tuning
- datasets
- allenai/scifact
---
# SciFact (Chat-Format Preparation)

This dataset is a chat-format preparation of SciFact for supervised fine-tuning (SFT).

## Format

This format is commonly referred to as:

- chat-format SFT data
- instruction-tuning conversations
- OpenAI-style `messages` format

## Included files

- `train.jsonl`
- `validation.jsonl`
- `stats.json`
- `prepare_scifact_unsloth.py`

## Source

- Base dataset: `allenai/scifact`

## Original Dataset Highlights

- Original dataset: `allenai/scifact`
- Focus: scientific claim verification with rationales.
- Labels: `SUPPORT`, `CONTRADICT`, and `NOT ENOUGH INFO` style supervision.
- Reported scale on source card: ~1.4K claims with evidence-containing abstracts.
- Paper: [Fact or Fiction: Verifying Scientific Claims](https://aclanthology.org/2020.emnlp-main.609/)

## Preparation summary

- Claim verification labels:
  - `SUPPORTS`
  - `CONTRADICTS`
  - `NOT ENOUGH INFORMATION`
- Abstract sentences are indexed (`[0]`, `[1]`, ...).
- One row is emitted per claim-evidence set.
- NEI rows are created when no support/contradict evidence is present.

Assistant response format:

- `Label: ...`
- `Rationale sentence ids: [...]`
- `Explanation: ...`

## Schema

Each JSONL row contains:

- `messages`
  - `user`: instruction + claim + title + indexed abstract
  - `assistant`: structured label/rationale/explanation
- `meta`: claim/doc ids, label, rationale sentence ids, split, variant

## Reproduction

```bash
python prepare_scifact_unsloth.py
```