trjxter/Kimi-K2.6-Technical-Reasoning-AddOn-3300x

Kimi-K2.6-Technical-Reasoning-AddOn-3300x

<p align="center"> <img src="artifacts/Kimi-K2.6-Distill-Addon.png" alt="Kimi-K2.6-Distill-Addon.png" width="100%"> </p>

This dataset is a technical reasoning add-on dataset generated with Kimi K2.6 as the teacher model.

The dataset was designed as an additional technical reasoning trace set for downstream SFT experiments, especially around math, graduate-level science, coding, and debugging/code-repair style prompts.

Dataset Summary

• —Dataset name: Kimi-K2.6-Technical-Reasoning-AddOn-3300x
• —Teacher model: Kimi-K2.6
• —Backend: W&B Inference
• —Target size: ~3,500 seed prompts
• —Final cleaned rows: 3301
• —Generation max output tokens: 32768
• —Temperature: 0.6
• —Top-p: 0.95

The original generation produced 3,414 completed rows. A cleaning pass removed malformed/meta-format artifacts, short invalid generations, and final-answer formatting issues, producing the final 3301 clean examples.

Domain Split

Source Seed Mix

Seed prompts were sampled and formatted from technical reasoning datasets, then answered by Kimi K2.6.

Format

Each row follows this schema:

{
  "id": "kimi_k2_6_technical_reasoning_addon_00000",
  "conversations": [
    {
      "role": "user",
      "content": "..."
    },
    {
      "role": "assistant",
      "content": "<think>...reasoning...</think>\n\nFinal answer: ..."
    }
  ],
  "input": "...",
  "output": "<think>...reasoning...</think>\n\nFinal answer: ...",
  "domain": "math",
  "source": "...",
  "seed_id": "...",
  "metadata": {...}
}

The assistant output uses a reasoning-trace format:

<think>
Reasoning trace here.
</think>

Final answer here.

Domains

Math

Competition-style and structured math prompts requiring derivations, exact answers, and clear final solutions.

Science

Graduate-level and general science prompts requiring careful conceptual reasoning and final answer selection.

Coding

Programming and coding-instruction prompts requiring solution reasoning, code generation, edge case analysis, complexity discussion, and final code where relevant.

Debugging

Coding/debugging-style prompts requiring diagnosis, repair-style reasoning, and final corrected solutions where relevant.

Generation Notes

The generation configuration used:

teacher_model = Kimi-K2.6
model_name = moonshotai/Kimi-K2.6
max_output_tokens = 32768
temperature = 0.6
top_p = 0.95

A small number of generations were rejected during cleaning if they contained meta-format artifacts, referenced dataset creation/system prompt details, had reasoning traces that were too short, or had incomplete final-answer formatting.

Intended Use

This dataset is intended for:

• —supervised fine-tuning
• —technical reasoning model training
• —math/science/coding reasoning experiments
• —reasoning trace distillation
• —add-on training data for larger technical reasoning datasets

Limitations

This dataset is synthetic and teacher-generated. It may contain mistakes, imperfect reasoning, or over-detailed traces. Users should evaluate model behavior carefully after training.

The dataset is an add-on set, not a full general assistant dataset. It is focused on technical reasoning domains.

License

Released under the Apache 2.0 license.