Claude Opus 4.6/4.7 Fine-Tuning Dataset

A synthetic fine-tuning dataset generated with Claude Opus 4.6/4.7 has been uploaded to Hugging Face.

• A total of 8,706 examples, an estimated 17,013,533 tokens, averaging 1,954 tokens/example
• Multi-turn accounts for 3,454 (39.7%), single-turn for 5,252 (60.3%)
• Split into full_train.jsonl (all 28 categories), instruct_train.jsonl (24 instructional categories), roleplay_train.jsonl (4 creative categories), and code_train.jsonl (coding+math)
• The top categories are coding (1,628), humanities (862), and science (737), in that order
• It states that only basic cleaning was applied, and that the data was processed in a way that suppresses refusal and safety responses

This is a synthetic fine-tuning dataset worth checking out, broadly covering coding, math, science, humanities, and creative writing.