SKU SFT-2403 · Sold by External
Unnatural Instructions
Product specifications
| SKU | SFT-2403 |
|---|---|
| Data type | SFT dataset |
| Volume | 68,478 instruction-input-output triplets (core set; 240,670 examples in the full set with paraphrases) |
| Size on disk | core_data.zip ~11.4 MB, full_data.zip ~18.3 MB (uncompressed core_data.jsonl ~61 MB) |
| Format | JSONL |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license (MIT) |
| Quality score | — |
| License | MIT License |
Unnatural Instructions is a model-generated instruction-following (SFT) dataset introduced by Honovich, Scialom, Levy and Schick (Tel Aviv University & Meta AI). The core set contains 68,478 instruction-input-output triplets (66,010 unique instruction lines in core_data.jsonl); the full set expands these to 240,670 examples by prompting the model to rephrase each instruction. Data was collected by prompting OpenAI's text-davinci-002 with three seed examples to elicit a fourth instruction, then generating the corresponding input, constraints and output. Each record has an `instruction`, and one or more `instances` carrying `input`, `instruction_with_input`, `constraints`, and `output`. The paper reports that despite containing a fair amount of noise, fine-tuning on this data rivals or surpasses models trained on manually curated datasets such as T0++ and Tk-Instruct. Because outputs are model-generated, they are used directly as SFT targets with no execution-based or human verification.