SKU SFT-2301 · Sold by External
Magicoder OSS-Instruct 75K & Evol-Instruct 110K
Product specifications
| SKU | SFT-2301 |
|---|---|
| Data type | SFT dataset |
| Volume | 186380 rows |
| Size on disk | 203 MB (OSS-Instruct) + 255 MB (Evol-Instruct) |
| Format | Parquet (Hugging Face) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | MIT (OSS-Instruct) / Apache-2.0 (Evol-Instruct) — generated with OpenAI models, subject to OpenAI terms |
Two companion supervised fine-tuning datasets released by ISE-UIUC with the Magicoder paper (arXiv:2312.02120). Magicoder-OSS-Instruct-75K holds 75,197 code problem/solution pairs across 9 programming languages, generated by GPT-3.5-turbo-1106 seeded from real open-source code snippets (OSS-Instruct). Magicoder-Evol-Instruct-110K holds 111,183 instruction/response pairs, a BigCode-decontaminated version of evol-codealpaca-v1. Neither contains tool/function-call traces; both are pure code-generation SFT data. Note the two sub-datasets carry different licenses (MIT vs Apache-2.0) and were generated with OpenAI models (subject to OpenAI usage terms).