Claude‑Opus 4.6 Trace Inversion: 9K High‑Fidelity Reasoning Traces

Claude‑Opus‑4.6‑TraceInversion‑9000x is a synthetic, multilingual chain‑of‑thought (CoT) dataset created by Jackrong using the novel Trace Inversion technique. The dataset contains 9,000 gzipped JSON‑L lines, each derived from Claude Opus 4.6’s "reasoning bubbles" – concise answer‑plus‑summary pairs – and transformed into full, step‑by‑step reasoning traces via the Jackrong/Trace‑Inverter‑4B model. Built on the Roman1111111/claude‑opus‑4.6‑10000x source, the dataset spans six languages (English, Chinese, Korean, Japanese, Russian, Spanish) and is licensed under Apache‑2.0.

The core contribution is the reconstruction of learnable CoT signals from compressed reasoning summaries, a process grounded in the information‑theoretic concept of negentropy. By enforcing the constraints of the original final answer and its summary, the Trace‑Inverter‑4B model synthesizes high‑quality synthetic CoT that preserves the style of Claude Opus while providing richer intermediate steps. This enables effective fine‑tuning (SFT) and direct preference optimization (DPO) for open‑source models such as Qwen and Unsloth, aiming to close the reasoning gap with proprietary systems.

Intended for text‑generation and reasoning tasks, the dataset is ideal for training and evaluating models on complex problem domains like mathematics, calculus, and logical puzzles. Its multilingual nature allows researchers to explore cross‑lingual reasoning capabilities, while the synthetic origin ensures scalability without manual annotation. The community has highlighted the dataset for its potential to boost reasoning in smaller models and for showcasing a reproducible method to generate learnable CoT data from any black‑box LLM that provides only final answers and brief summaries.