{"id":40821,"date":"2026-05-05T15:27:03","date_gmt":"2026-05-05T13:27:03","guid":{"rendered":"https:\/\/www.graviton.at\/letterswaplibrary\/open-source-tool-for-generating-and-cleaning-synthetic-instruction-tuning-datasets\/"},"modified":"2026-05-05T15:27:03","modified_gmt":"2026-05-05T13:27:03","slug":"open-source-tool-for-generating-and-cleaning-synthetic-instruction-tuning-datasets","status":"publish","type":"post","link":"https:\/\/www.graviton.at\/letterswaplibrary\/open-source-tool-for-generating-and-cleaning-synthetic-instruction-tuning-datasets\/","title":{"rendered":"Open Source Tool For Generating And Cleaning Synthetic Instruction-tuning Datasets"},"content":{"rendered":"

<\/p>\n

\n

Built this because I wanted a reproducible way to build fine-tuning datasets without doing it all by hand.<\/p>\n

You give it seed prompts or an existing dataset, it generates instruction-output pairs via any OpenRouter model, scores them with a local or remote LLM judge, and exports a clean JSONL you can use directly for training. <\/p>\n

You can also ingest datasets straight from HuggingFace and filter or relabel them through the same pipeline.<\/p>\n

The export step lets you set a score threshold and a train\/val split ratio so what comes out is ready to use.<\/p>\n

MIT licensed, everything is stored locally, no data leaves your machine unless you choose a cloud judge backend.<\/p>\n

Github project link is in comments below \ud83d\udc47<\/p>\n<\/div>\n

submitted by \/u\/gvij <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>