Tivaphraen/Qwen3.7_5k_fr60_fixed

Qwen 3.7 Max Thinking — Distilled Reasoning Dataset (FR60, cleaned)

5,000 chain-of-thought (CoT) reasoning traces, ~60% machine-translated to French, derived from the original dataset `WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K`.

Each example contains a problem, a detailed step-by-step reasoning trace (in the Qwen 3.7 Max Thinking style), and a concise final answer.

Source and translation

This dataset is a partial translation of the original English dataset `Qwen3.7_Max_Thinking_dataset_5K` published by WithinUsAI. About 60% of the 5,000 examples were machine-translated into French, the remaining ~40% were kept as-is in the original English — hence the fr60 suffix.

The translation was performed with a script using the [tencent/Hy-MT2-1.8B-GGUF](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF) model, quantized Q6, run locally.

The original dataset was generated programmatically (generate_dataset.py) using templated generators with randomized parameters across 10 reasoning domains, then deduplicated via SHA-256 hashing (both at the entry and problem level).

Format

JSONL, one conversation per line, messages structure (chat/SFT-compatible):

{"messages": [
  {"role": "user", "content": "Résolvez l'équation quadratique : 2x^2 + 8x + 6 = 0"},
  {"role": "assistant", "content": "Résoudre 2x^2 + 8x + 6 = 0.\nEn utilisant la formule quadratique : x = [-b +/- sqrt(b^2 - 4ac)] / (2a)\n...\n\nx = -1, x = -3"}
]}

The response style consistently follows the same pattern: restate the problem → recall the relevant formula → step-by-step computation → self-verification → isolated final answer.

Stats

Category breakdown (from the original dataset)

Known issue — fixed in this version

The original dataset contained a generation bug affecting ~39% of quadratic-equation examples (~104 of 267, ~2% of the full dataset): the intermediate root-computation step displayed a spurious duplicated value, e.g.

x1 = [-11 + 5] / 2 = -3 = -8
x2 = [-11 - 5] / 2 = -8 = -3

The final answer was always correct (x = -8, x = -3); only the intermediate reasoning step contained this erroneous duplication. This version has been cleaned: the spurious duplicate value has been removed from all affected examples, keeping only the mathematically correct value. No examples were dropped — all 5,000 rows are preserved.

Use cases

• —Knowledge distillation
• —Supervised fine-tuning (SFT) for CoT reasoning in French/English
• —Research on structured/bilingual reasoning formats

License

The license of the original dataset is not specified by the author (WithinUsAI). Please check with the source before any commercial use, and keep the attribution below.

Citation

If you use this dataset, please cite the original dataset it is derived from:

@misc{qwen37maxthinking5k,
  title        = {Qwen 3.7 Max Thinking --- Distilled Reasoning Dataset},
  author       = {WithinUsAI},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K}}
}

This dataset (fr60, cleaned version) is a partial French translation and cleanup of the above original dataset. Translation performed with tencent/Hy-MT2-1.8B-GGUF (Q6 quantization). Please credit both the original dataset and this derived version if you reuse it.