SKU SFT-2402 · Sold by External
Self-Instruct
Product specifications
| SKU | SFT-2402 |
|---|---|
| Data type | SFT dataset |
| Volume | 82612 instances (prompt–completion pairs) in the self_instruct config |
| Size on disk | 20.5 MB (self_instruct config dataset_size = 20,527,462 bytes; download_size ≈ 24.1 MB per HF dataset_info) |
| Format | Text; HF configs served as prompt/completion string pairs (Parquet via datasets-server); original repo ships JSON/JSONL |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (HF) |
| Quality score | — |
| License | Apache License 2.0 |
Self-Instruct is a synthetic instruction-tuning dataset produced by the framework of the same name, which bootstraps a language model's own generations from a small pool of 175 human-written seed tasks (1 instruction + 1 instance each). Applying the pipeline to the vanilla GPT-3 (text-davinci-001), the authors generate instructions, inputs, and outputs, then filter invalid or near-duplicate items (ROUGE-L similarity threshold) to yield, per the paper's Table 1, 52,445 instructions and 82,439 instances. The Hugging Face release exposes the primary machine-generated data as the self_instruct config with 82,612 prompt/completion pairs, alongside a 252-example expert-written human_eval (user-oriented) split and re-formatted super_natural_instructions and P3/prompt_source configs. Records are used directly as supervised fine-tuning demonstrations; there is no automatic correctness grader for the generated split. The data and code are released under Apache-2.0, though the generated split was produced with OpenAI models (text-davinci-001), so downstream users should review OpenAI usage terms. The work is by Yizhong Wang and collaborators (University of Washington and the Allen Institute for AI, with co-authors at Tehran Polytechnic, Arizona State University, and Johns Hopkins University).