Ended up with some tokens to burn on a Claude Max plan. Assembly began during 4.6 and moved to 4.7. Model is tagged. The development evolved as it went along. The dataset has not been manually reviewed. It's entirely Claude developed.
The reasoning is not Claude's actual chain-of-thought (cot) and is not summarized cot. It's a fully synthetic cot created as part of the Assistant response to mimic the type of "thinking"… See the full description on the dataset page:
https://huggingface.co/datasets/ansulev/opus-4.6-4.7-reasoning-8.7k.
