xtools-at/Claude-Haiku-5.5-High-Reasoning-TraceInversion-3000x

⚗️ Claude Haiku 5.5 High Reasoning - Synthetic Reasoning via Trace Inversion

This dataset explores frontier reasoning distillation with Claude Haiku 5.5 in high reasoning mode, using trace inversion. The reasoning questions were generated with Gemini 3.6 Flash in high reasoning mode across 10 categories covering advanced mathematics, theoretical computer science, formal logic, decision theory, scientific deduction, and related reasoning-intensive domains.

Trace inversion takes a model's visible answer and a short reasoning summary, then uses a separate model to expand those clues into a longer, plausible synthetic reasoning trace. The resulting dataset is intended to explore reasoning-data distillation through synthetic reconstruction of reasoning summaries. The generated traces should not be interpreted as recovered internal reasoning or as faithful representations of Claude's internal process.

🧵 How it was made

1. Prompt generation: A total of 4,000 reasoning questions were generated with Gemini 3.6 Flash in high reasoning mode. The questions are equally distributed across 10 reasoning categories, covering mathematical, computational, philosophical, strategic, statistical, systemic, scientific, and normative reasoning.
2. Response collection: The prompts were sent to Claude Haiku 5.5 in high reasoning mode. The responses were saved, along with reasoning summaries where they were available.
3. Initial filtering: After response collection, answers longer than 24,000 characters were filtered out, leaving 3,577 entries eligible for trace inversion.
4. Trace inversion: Responses with an available reasoning summary were processed with xtools-at/Qwen3.5-9B-Trace-Inverter. The summaries were expanded into full synthetic reasoning traces using the associated prompt and answer as context.
5. Final filtering: Another 496 answers were dropped because their thinking traces exceeded 8,000 tokens. An additional 2 entries were removed due to the 24,000-character limit. This resulted in a final dataset of 3,079 entries.

📦 What is included

• 4,000 reasoning questions generated with Gemini 3.6 Flash.
• 10 reasoning categories, equally represented with 400 questions per category.
• Saved Claude Haiku 5.5 responses generated in high reasoning mode.
• Reasoning summaries for responses where a summary was returned.
• Synthetic trace-inversion outputs generated from available reasoning summaries using xtools-at/Qwen3.5-9B-Trace-Inverter.
• 3,079 filtered entries containing the final dataset after the response-length and reasoning-trace filtering stages.

Reasoning categories

The original 4,000 questions are equally composed of the following categories:

1. Advanced Pure Mathematics
2. Mathematical Logic & Foundations
3. Theoretical Computer Science & Complexity
4. Applied Mathematics, Optimization & Reasoning
5. Epistemology & Paradoxes
6. Game Theory & Decision Theory
7. Causal Inference & Statistical Logic
8. Systems Thinking & Complex Diagnostics
9. Scientific Methodology & Deduction
10. Legal, Dialectical & Normative Logic

Each category contains ~400 questions, providing an even distribution across the dataset's intended reasoning domains.

🗂️ Available dataset splits

The dataset is available in four split configurations, providing access to the complete collection or to separate training, evaluation, and test subsets.

Split	Number of entries	Share of dataset
all	3,079	100%
train	2,799	90%
eval	140	5%
test	140	5%
Total	3,079	100%

• all — The complete dataset containing all 3,079 entries.
• train — The training subset containing 2,799 entries, intended for supervised fine-tuning.
• eval — The evaluation subset containing 140 entries, intended for monitoring training progress and evaluating intermediate model performance.
• test — The test subset containing 140 entries, intended for final evaluation of trained models.

The train, eval, and test subsets account for all entries, with an approximately 90% / 5% / 5% distribution. The all split provides access to the complete collection rather than an additional set of unique examples.

For reliable evaluation, the evaluation and test subsets should be kept separate from training data during fine-tuning.

📋 Example data

Each example contains the original prompt, the model response, and the available reasoning-related fields, where available. The synthetic reasoning was generated through trace inversion using the available reasoning summaries.

A shortened example, illustrating the general structure, follows:

{
  "source": "gemini3.6-flash",
  "model": "anthropic/claude-haiku-5.5",
  "messages": [
    {
      "role": "user",
      "content": "Convert 1994 to Roman numerals and walk me through how you got it."
    },
    {
      "role": "assistant",
      "reasoning": "Synthetic reasoning generated through trace inversion from the available reasoning summary.",
      "content": "1994 = MCMXCIV. 1000 = M, 900 = CM, 90 = XC, and 4 = IV.",
      "reasoning_summary": "Explains the Roman numeral breakdown, including subtractive notation such as CM (1000 − 100) and IV (5 − 1)."
    }
  ]
}

Note: The example above illustrates the general schema and is not intended to represent a verified record from the dataset. The reasoning and summary fields are illustrative.

Field descriptions

• source — Identifies the source of the original prompts. In this dataset, the questions were generated with Gemini 3.6 Flash.
• model — Identifies the model used to generate the responses: Claude Haiku 5.5 in high reasoning mode.
• reasoning — Contains synthetic reasoning generated by applying xtools-at/Qwen3.5-9B-Trace-Inverter to the available reasoning summary, using the associated prompt and answer as context. These traces are synthetic reconstructions, not recovered original internal reasoning.
• content — The final response generated by Claude Haiku 5.5 in high reasoning mode.
• reasoning_summary — The reasoning summary returned alongside the model response, where available, before trace-inversion processing.

💶 Cost Breakdown

The dataset was produced using model-generated questions, Claude Haiku 5.5 responses, and synthetic reasoning generation through trace inversion. The total cost depends on the API usage, token counts, inference setup, and compute requirements of each component.

Component	Approximate Cost
Gemini 3.6 Flash question generation x4,000	$0.00 ($5/mo subscription)
Claude Haiku 5.5 high-reasoning responses x4,000	$14.50
Trace-Inverter inference x3,577	$3.20
Total	$17.70
Average per entry (x3,079)	0.00575

🎯 Intended use

This dataset is intended for research into reasoning-data distillation, supervised fine-tuning, and the use of synthetic reasoning traces as training data. It combines 3,000+ reasoning questions across 10 different categories with Claude Haiku 5.5 responses generated in high reasoning mode.

By applying trace inversion to responses with available reasoning summaries, the dataset explores the use of synthetic reasoning reconstruction as a method for constructing reasoning-focused distillation data. The resulting examples may be useful for studying how synthetic reasoning affects downstream model training, reasoning performance, and generalization across domains.

The expanded traces are model-produced reconstructions, not ground-truth reasoning or recovered internal processes. The dataset should be treated as an experimental research resource rather than a validated reasoning benchmark.

⚠️ Limitations

• Synthetic, not recovered: The reasoning traces were generated using xtools-at/Qwen3.5-9B-Trace-Inverter. They are not Claude Haiku 5.5's original hidden reasoning.
• May rationalize an answer: A trace inverter can produce a coherent-looking explanation for an answer that is mistaken, incomplete, or unsupported. The generated traces should not be treated as proof of correctness.
• Generated questions: The prompts were generated with Gemini 3.6 Flash and may contain ambiguities, errors, unintended assumptions, or uneven difficulty within and across categories.
• Filtering and selection bias: The dataset excludes answers longer than 24,000 characters and thinking traces exceeding 8,000 tokens. This may introduce selection bias, as particularly long or complex reasoning examples are not represented in the final dataset.
• Missing reasoning summaries: Not every response necessarily contains a reasoning summary. Only examples with available reasoning summaries can be used for trace inversion, and entries without them are not supplemented with independently generated reasoning.
• High-reasoning mode does not guarantee correctness: The use of Claude Haiku 5.5 in high reasoning mode does not ensure that every answer is accurate, rigorous, or complete.
• Synthetic-data bias: The dataset reflects the behavior and limitations of the question-generation model, the response model, and the trace-inversion model. These characteristics may be reinforced during downstream fine-tuning.
• Not a benchmark: The dataset was assembled for reasoning-data generation and distillation research. Its category balance does not establish equivalent difficulty, quality, or coverage across all reasoning domains.

📚 Sources and related work

• Question generation model: Gemini 3.6 Flash.
• Response model: Claude Haiku 5.5 in high reasoning mode.
• Trace inverter: xtools-at/Qwen3.5-9B-Trace-Inverter.
• Background paper: How to Steal Reasoning Without Reasoning Traces by Tingwei Zhang, John X. Morris, and Vitaly Shmatikov (2026).

⚖️ Licensing and use

The dataset contains questions generated with Gemini 3.6 Flash, responses generated by Claude Haiku 5.5 through its high reasoning mode, and synthetic reasoning outputs produced using the Qwen3.5-9B-Trace-Inverter model.

The applicable licensing and usage terms may differ across these components. The licenses of the question-generation model, response model, and reasoning-generation model should not be assumed to grant identical rights over the combined dataset. Review the relevant model and provider terms, as well as the applicable terms for the trace-inversion model, before redistribution or commercial use.

The metadata is marked other pending an explicit license determination for the combined dataset.

📝 Citation

If you use this dataset, please cite the dataset repository and the trace-inversion paper. Paper reference:

@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}
}