---
language:
  - en
license: apache-2.0
task_categories:
  - conversational
  - text-generation
tags:
  - reasoning
  - distillation
  - sft
  - deepseek
  - kimi
  - qwen
size_categories:
  - 10K<n<100K
---

# distill-sft-26k

A merged SFT dataset of **25,973** high-quality reasoning conversations distilled from three frontier models.

## Source Datasets

| Source | Rows | Reasoning | Domain |
|--------|------|-----------|--------|
| Spakie/DeepSeek-V4-Pro-distilled | 17,670 | No | General |
| trjxter/Kimi-K2.6-Reasoning-3300x-WandB | 3,303 | Yes (all, ``) | General / Science / Math |
| WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K | 5,000 | Yes (all) | Math / Science |

## Format

Standard chat format with optional reasoning:

```json
{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "...", "reasoning_content": "..."}
  ],
  "source": "deepseek_v4_pro | kimi_k2_6 | qwen3_7_max",
  "domain": "general | science | math",
  "id": "..."
}
```

- `reasoning_content` present on 8,303 rows (32%)
- Shuffled with seed=42

## Loading

```python
from datasets import load_dataset
ds = load_dataset("GenueAI/distill-sft-26k", split="train")
```

## License

Apache 2.0. See source dataset licenses for details.