xtools-at/trace-inversion-opus-synthetic

🫧 Trace Inversion - Synthetic Reasoning Dataset

💡 Overview

This dataset is designed to help train language models to reconstruct detailed reasoning traces from compressed reasoning summaries, a process known as trace inversion.

Each example pairs a problem and its known final answer with one or more short reasoning summaries, then provides a more detailed reconstructed reasoning trace as the training target. The goal is to teach models to expand compact reasoning information into coherent, structured explanations.

The dataset combines two collections of synthetic reasoning examples derived from Claude Opus 4.6 and Claude Opus 4.7 outputs, and was used to train xtools-at/Qwen3.5-4B-Trace-Inverter.

📚 Data sources

The dataset incorporates all entries from the following parent datasets, subject to a maximum-length filter of 8,000 tokens:

• Roman1111111/claude-opus-4.6-10000x, through Jackrong/Claude-opus-4.6-TraceInversion-9000x.
• lordx64/reasoning-distill-claude-opus-4-7-max, through Jackrong/Claude-opus-4.7-TraceInversion-5000x.

The parent collections provide the original problems, answers, and compressed reasoning information. The derived trace-inversion collections provide the corresponding reconstructed reasoning traces.

🧩 Dataset structure

The dataset uses OpenAI chat-style conversations serialized as JSONL, with one example per line. Each record contains a messages array with three roles:

• system — Defines the trace-inversion task and the rules for reconstructing a reasoning trace.
• user — Contains the original problem, the model's final answer, compressed reasoning bubbles, and an instruction to reconstruct the full trace.
• assistant — Contains the target response: a detailed, coherent reconstructed trace wrapped in <think>...</think>.

💬 Simplified example

A typical example follows this structure:

{
  "messages": [
    {
      "role": "system",
      "content": "Reconstruct a detailed reasoning trace from the problem, final answer, and reasoning bubbles. Output only the trace inside <think>...</think>."
    },
    {
      "role": "user",
      "content": "Problem: Lyka needs $120 more to buy a $160 smartphone. She has $40 and plans to save over 8 weeks.\n\nModel's final answer: $15 per week.\n\nReasoning Bubbles:\nRemaining amount: $160 - $40 = $120\nWeekly savings: $120 / 8 = $15\n\nReconstruct the full reasoning trace."
    },
    {
      "role": "assistant",
      "content": "<think>Lyka needs $160 - $40 = $120 more. Saving this amount over 8 weeks means she should save $120 / 8 = $15 per week.</think>"
    }
  ]
}

This example is abbreviated for readability; actual records contain the full task instructions and reconstructed reasoning traces.

📊 Dataset composition

Split	Examples	File
Training	11,103	train_trace_clean_l4k/train.jsonl
Evaluation	555	train_trace_clean_l4k/eval.jsonl
Total	11,658

Both splits follow the same message structure. The training set is intended for supervised fine-tuning, while the separate evaluation set supports measuring performance on held-out examples.

🔧 Processing

• All entries from both parent datasets are included, subject to the length filter.
• Examples exceeding 4,000 characters are filtered out.
• No additional sampling or intentional downsampling is applied.

The character limit keeps examples manageable for training while retaining a broad range of reasoning tasks and trace lengths.

🎯 Intended uses

• Training smaller language models for trace inversion.
• Research into reasoning distillation and synthetic reasoning data.
• Supervised fine-tuning with compressed reasoning summaries as input and reconstructed traces as targets.
• Exploring how models can expand concise reasoning representations into more detailed explanations.

⚠️ Limitations

The reconstructed traces are synthetic outputs, not verified records of a source model's private internal reasoning. They may contain mistakes, unsupported steps, or explanations that rationalize an incorrect final answer.

A known answer does not guarantee that the accompanying reasoning is correct. Users should evaluate trace quality and task performance independently before using the data in downstream applications.

The dataset's language coverage and domain distribution reflect its upstream sources; per-language and per-domain proportions have not been established here.

⚖️ Licensing

The upstream collections have their own provenance and usage terms. Users should review the licenses and applicable policies of the parent datasets and source models before redistribution or commercial use.

🔗 References

• xtools-at/Qwen3.5-4B-Trace-Inverter
• Jackrong/Trace-Inverter-4B
• amkkk/Trace-Inverter-4B-NoBubble
• Jackrong/Claude-opus-4.6-TraceInversion-9000x
• Jackrong/Claude-opus-4.7-TraceInversion-5000x
• Roman1111111/claude-opus-4.6-10000x
• lordx64/reasoning-distill-claude-opus-4-7-max
• Zhang, Morris, and Shmatikov, How to Steal Reasoning Without Reasoning Traces, 2026.

@misc{zhang2026stealreasoning,
  title         = {How to Steal Reasoning Without Reasoning Traces},
  author        = {Tingwei Zhang and John X. Morris and Vitaly Shmatikov},
  year          = {2026},
  eprint        = {2603.07267},
  archivePrefix = {arXiv}
}