khazarai/kimi-2.5-high-reasoning-250x

---
license: apache-2.0
language:
- en
pretty_name: distilled dataset
task_categories:
- text-generation
tags:
- reasoning
- distillation
size_categories:
- n<1K
---

# kimi-2.5-high-reasoning-250x

kimi-2.5-high-reasoning-250x is a long-form reasoning distillation dataset designed to train smaller language models to perform deep analytical thinking and structured multi-step reasoning.

The dataset was generated using **Kimi-2.5-thinking** as a teacher model to produce detailed reasoning traces and final answers for complex questions across multiple technical, scientific, historical, and strategic domains.

The primary goal of the dataset is knowledge distillation: transferring reasoning patterns from a powerful teacher model to smaller models in the 0.6B–14B parameter range.

Each dataset row contains:

- A complex reasoning question
- A long structured reasoning trace
- A clear final answer summarizing the reasoning outcome

The dataset is optimized for reasoning-focused supervised fine-tuning (SFT).

## Dataset Statistics

| Property | Value |
| ----------| ----- |
| Dataset Name | kimi-2.5-high-reasoning-250x |
| Total Samples | 250 |
| Total Tokens | 1,114,407 |
| Average Tokens per Sample | ~4,457 |
| Max_seq_length | 8000 |
| Format | JSON |
| Teacher Model | Kimi-2.5-Thinking |
| Dataset Type | Synthetic reasoning distillation |

**Question:**

A complex problem requiring deep reasoning, often involving multi-disciplinary knowledge.

**Model_thought:**

A detailed reasoning trace including:
- assumptions
- intermediate conclusions
- comparisons
- logical deductions
- analytical steps

These reasoning traces are intentionally long to help smaller models learn structured thinking patterns.

**Model_response:**

A concise final answer summarizing the reasoning outcome.

### Covered Domains

The dataset spans 32 reasoning domains across coding, math, history, medicine, philosophy, and strategic analysis.