Why this matters
Most commercial LLM outputs expose only compressed summaries (“reasoning bubbles”), which lack intermediate steps and create an information gap when used as direct supervision for smaller models. This dataset fills that gap by reverse-engineering detailed, step-by-step reasoning traces from Claude Opus 4.6 summaries under the constraints of original problem + final answer + bubble, producing token-level CoT supervision that is more learnable for student models.
What Sets It Apart
- Reconstructed CoT at scale — 9,000 curated examples where each sample pairs the original problem and final answer with a synthesized, logically complete chain-of-thought. So what: provides continuous token-level supervision rather than short summaries, improving the signal for reasoning distillation.
- Trace-inversion methodology — built from Roman1111111/claude-opus-4.6-10000x and generated by Jackrong/Trace-Inverter-4B using a negentropy-guided inversion pipeline. So what: aims to preserve style and depth of commercial outputs while restoring missing intermediate steps.
- Training-ready packaging — exported as gzip-compressed .jsonl (SFT format), includes merged responses with
<think>...</think>blocks and labeled fields for direct use with SFT/DPO workflows. So what: reduces engineering friction for researchers wanting to run supervised fine-tuning or preference optimization experiments. - Multilingual & cross-domain examples — contains English plus other languages and cases across math, coding, physics, and logical reasoning. So what: useful for both instruction-following and reasoning-focused benchmarks across languages.
Who it’s for — and tradeoffs
Great fit if you are experimenting with reasoning distillation, SFT or DPO fine-tuning of mid-sized open models (e.g., Qwen, LLaMA forks) and need higher-fidelity CoT supervision than short summaries provide. Look elsewhere if you require human-verified ground-truth traces (these are synthetic reconstructions, not internal model internals) or if your workflow cannot tolerate any downstream label noise.
Practical notes: the dataset is under Apache-2.0, was packaged on May 19, 2026, and includes a disclaimer that reconstructed traces may rationalize subtle upstream errors. For numeric or code-heavy samples we recommend automated unit tests or execution-based verification before using them as high-stakes supervision signals.
