Conversational QA pairs describing Qwen3-8B chain-of-thought reasoning traces. Generated by prompting Gemini 2.0 Flash to describe observable facts about CoT text with zero logical leaps.
Training data for activation oracles — models that read their own activations and answer questions about their reasoning. The oracle sees strided activations at sentence boundaries, not the CoT text itself.
The training signal: simple question → factual… See the full description on the dataset page:
https://huggingface.co/datasets/ceselder/cot-conversational-qa-v1.
