PXIN/reasoning-chatml-3k

---
language:
- en
license: apache-2.0
tags:
- reasoning
- chatml
- thinking
- fine-tuning
- summarized-reasoning
size_categories:
- 1K<n<10K
---

# Summarized Reasoning ChatML 3K

**3,131 summarized reasoning examples** in ChatML format with extended thinking (`<think>` tags), cleaned of refusals and ready for fine-tuning.

## Technical Note: "Summarized" vs. "Raw" CoT
Following the release of Claude 4, Anthropic standardized on "Extended Thinking" outputs that represent a highly sophisticated **summary** of the model's internal Chain-of-Thought (CoT) rather than the raw, unedited internal process. This dataset distills those high-level reasoning summaries to teach small models the articulation and logical structure of frontier models like Claude 4.6 Opus.

## Sources
Combined and cleaned from three summarized reasoning distillation datasets:

| Source | Original Rows | After Cleaning | Description |
|--------|--------------|----------------|-------------|
| [nohurry/Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) | 2,326 | 2,249 | Claude Opus 4.6 summarized reasoning distillation |
| [Jackrong/Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) | 633 | 632 | Qwen 3.5 extended thinking |
| [TeichAI/claude-4.5-opus-high-reasoning-250x](https://huggingface.co/datasets/TeichAI/claude-4.5-opus-high-reasoning-250x) | 250 | 250 | Claude Opus high-quality reasoning examples |

## What was done
1. **Format unification**: All sources converted to ChatML with `<think>` tags.
2. **Refusal removal**: 60 true model refusals purged via a precise pattern-matching audit.
3. **Consolidation**: 3,131 high-quality rows merged into a single JSONL file.

## Intended Use
Fine-tuning small language models (0.5B-9B) to improve logical articulation and structured thinking. By training on these summaries, small models learn to present reasoning with the clarity and depth of frontier architectures.