Why this matters
Large instruction‑tuning collections typically show what an assistant should answer; this dataset instead foregrounds how an assistant "thinks" by pairing each assistant response with a long, synthetic <think> block. If your goal is to expose a model to deliberative answer structure, multi‑turn scaffolding, and domain‑diverse expert‑style responses, this dataset offers a compact, ready‑to‑consume set of examples derived from Claude Opus 4.6 and 4.7 waves.
What Sets It Apart
- Synthetic, per‑example reasoning: every assistant message includes a 150–500 word
<think>segment that models deliberation, alternatives, and trade‑offs rather than only producing the final answer. This is intended to teach reasoning style and structure rather than to provide ground‑truth cognitive traces. - Broad domain coverage with expert tone: 28 populated categories (coding, math, sciences, humanities, arts, finance, medicine, law, etc.), plus roleplay tracks that maintain character voices. The dataset mixes single‑turn and multi‑turn examples, with ~39.7% multi‑turn conversations to train follow‑ups and context continuation.
- Compact, split for targeted tuning: four JSONL splits—full (8,706 examples), instruct (7,217), roleplay (1,489), and code (1,840)—so you can fine‑tune for general instruction behavior, coding/math, or characterized roleplay without importing unrelated examples.
Who It's For and Tradeoffs
Great fit if you: want to fine‑tune or benchmark an LLM to mimic deliberative answer structure; need synthetic chain‑of‑thought style examples for SFT experiments; or are researching prompt/response decomposition and multi‑turn consistency. The dataset is also useful for prompt‑engineering research because it surfaces model‑style reasoning and varied system prompts (5,814 unique system prompts).
Look elsewhere if you: require human‑verified ground truth — the dataset is not manually reviewed and contains fully synthetic reasoning blocks that can hallucinate or include unsafe content; need datasets that enforce refusals, safety hedging, or alignment signals — this collection intentionally excludes refusals and safety hedging to maximize capability examples.
Structure & Provenance
Key metrics: 8,706 examples (~17M estimated tokens), 28 categories, 100% examples with reasoning blocks, ~3,454 multi‑turn examples. Teacher sources are Claude Opus 4.6 (≈53.7% of examples) and Claude Opus 4.7 (≈46.3%). The dataset author is "angrygiraffe" and the package is hosted on Hugging Face under an Apache‑2.0 license. The readme emphasizes that the <think> blocks are synthetic assistant deliberations, not model internals or human annotations.
Practical notes
- Use splits to limit domain noise when fine‑tuning (e.g., use code_train.jsonl for coding/math models). - Treat
<think>blocks as a training signal for reasoning style, not as verifiable chain‑of‑thought evidence. - Because examples are not human‑reviewed and lack safety refusals, include downstream alignment/safety filters before deploying models fine‑tuned on this data.
