SKU SFT-2401 · Sold by External
Super-NaturalInstructions
Product specifications
| SKU | SFT-2401 |
|---|---|
| Data type | SFT dataset |
| Volume | 1,616 NLP tasks |
| Size on disk | ~8.15 GB (Hugging Face mirror, all splits: 7.22 GB train + 937 MB test + 2.75 MB validation) |
| Format | JSON task files (repo) / JSONL instance rows (HF mirror) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license (Apache-2.0) |
| Quality score | — |
| License | Apache License 2.0 |
Super-NaturalInstructions (NIv2) is a community-built benchmark of 1,616 diverse NLP tasks spanning 76 task types (classification, extraction, infilling, sequence tagging, rewriting, composition, question generation, etc.), released by the Allen Institute for AI and collaborators. Each task is a JSON file containing an expert-written Definition (the instruction), a set of Positive Examples and Negative Examples (each with input, output, and an explanation), and a large list of Instances (input/output pairs). The benchmark was designed to test whether models can follow natural-language instructions on tasks unseen during training, and it was used to train the Tk-Instruct instruction-following models. This listing verifies against the original allenai/natural-instructions GitHub repository (task files, Apache-2.0 LICENSE) and the widely used preprocessed Hugging Face mirror Muennighoff/natural-instructions, which flattens instances into rows with fields task_name, id, definition, inputs, targets. The HF mirror reports 6,164,188 train, 5,995 validation, and 982,285 test instance rows.