README.md

---
language:
- en
pretty_name: SciTLDR Chat-Format
license: apache-2.0
task_categories:
- summarization
task_ids:
- text2text-generation
source_datasets:
- allenai/scitldr
tags:
- text
- science
- summarization
- chat-format
- instruction-tuning
- datasets
- allenai/scitldr
- arxiv:2004.15011
---
# SciTLDR (Chat-Format Preparation)

This dataset is a chat-format preparation of SciTLDR for summarization 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_scitldr_unsloth.py`

## Source

- Base dataset: `allenai/scitldr`
- Variants used:
  - `A`
  - `AIC`
  - `FullText`

## Original Dataset Highlights

- Original dataset: `allenai/scitldr`
- Focus: extreme summarization of scientific papers (TLDR generation).
- Reported scale on source card: 5.4K TLDRs over ~3.2K papers.
- Multi-target setup: each paper can have multiple valid TLDR summaries.
- Paper: [TLDR: Extreme Summarization of Scientific Documents](https://arxiv.org/abs/2004.15011)

## Preparation summary

- Task: one-sentence scientific TLDR generation.
- User input is built from paper `title` and `source`.
- Assistant target is drawn from `target`.
- Supports:
  - `target-policy first`: first target only
  - `target-policy all`: one row per target
- Final train/validation splits are balanced across `A`, `AIC`, and `FullText`.

## Schema

Each JSONL row contains:

- `messages`
  - `user`: instruction + title + paper content
  - `assistant`: TLDR summary sentence
- `meta`: split, source variant, paper_id, target index/count

## Reproduction

```bash
python prepare_scitldr_unsloth.py --target-policy all
```