Why this matters
This dataset captures the internal extended-thinking traces and final answers produced by Anthropic's Claude Opus 4.7 for 8,124 prompts — explicitly prepared to help researchers and engineers supervise or distill reasoning and chain-of-thought behavior into open-source models.
What Sets It Apart
- Direct extended-thinking outputs: each row includes a separate
thinkingfield alongside theresponse, allowing supervised fine-tuning (SFT) approaches to learn from the model's intermediate reasoning as well as its final answer. - Provenance clarity: prompts were drawn from multiple source corpora but all responses and thinking fields are newly generated by
claude-opus-4-7, so the dataset is internally consistent for experiments that rely on a single high-quality generator. - Compact, structured metadata: rows include
messages(chat-format),systemprompts, token usage and inference geo fields — useful for filtering by cost/latency characteristics or for auditing generation context. - Practical packaging: ~8.1K examples in optimized Parquet format (train split), with an Apache‑2.0 package license for the dataset bundle (note: generated content remains subject to Anthropic's usage policies).
Who it's for — and tradeoffs
Great fit if you want supervised fine-tuning or distillation targets that expose chain-of-thought-style intermediate traces (e.g., SFT that concatenates thinking+answer, or student models trained to emulate intermediate steps). Also useful for prompt‑engineering research comparing prompt sources vs. model reasoning.
Look elsewhere if you need extremely large-scale corpora (this is ~8k examples), multilingual data (this dataset is English-only), or if your use case cannot accept content produced by a proprietary model without explicit compliance with Anthropic's terms. Be aware the dataset contains synthetic model outputs — potential factual errors, biases, or unsafe content from the generator should be audited before downstream use.
Where it fits
This dataset sits between small, hand-curated chain-of-thought collections and massive synthetic corpora: it’s compact but focused on complete Claude Opus 4.7 extended-thinking traces, making it convenient for controlled SFT/distillation experiments or for building evaluation sets that require both intermediate reasoning and final answers.
