bmeyer2025/glm5-reasoning-traces

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- distillation
- chain-of-thought
- glm-5
- math
- logic
- code
pretty_name: GLM-5 Reasoning Traces
size_categories:
- 1K<n<10K
---

# GLM-5 Reasoning Traces

<p align="center">
<img src="https://raw.githubusercontent.com/brianmeyer/distillreasoning/main/images/distillation_brain.png" alt="Large glowing neural network brain streaming reasoning chains into a compact crystal" width="600">
</p>

Reasoning traces generated by GLM-5 (744B MoE, 40B active parameters) via Ollama cloud. Designed for knowledge distillation — training smaller models to reason like frontier models.

## What's in this dataset

Each row contains a problem sent to GLM-5 and its full response including the `<think>` reasoning chain.

| Field | Description |
|-------|-------------|
| `id` | Unique problem identifier |
| `source` | Origin dataset: `gsm8k`, `math`, `arc`, `humaneval` |
| `problem` | The original question |
| `expected_answer` | Ground truth answer (where available) |
| `thinking` | GLM-5's full step-by-step reasoning chain |
| `response` | GLM-5's final answer |

## Problem sources

| Source | Count | Type |
|--------|-------|------|
| GSM8K | ~1,050 | Math word problems (numeric answers) |
| MATH | ~160 | Competition math (algebra, geometry, number theory) |
| ARC-Challenge | ~320 | Science reasoning (multiple choice) |
| HumanEval | ~130 | Coding problems (Python functions) |

## How it was generated

```python
import ollama

response = ollama.chat(
model="glm-5:cloud",
messages=[
{"role": "system", "content": "You are a reasoning expert. Think through each problem step by step in detail before giving your final answer. Show all your work."},
{"role": "user", "content": problem},
],
think=True, # Captures the full <think> chain
)
thinking = response.message.thinking
answer = response.message.content
```

Traces were filtered to keep only:
- Correct answers (verified for GSM8K and ARC)
- Thinking chains with 50+ tokens (model engaged deeply)
- Non-repetitive reasoning (no looping)

## Related

- **SFT-formatted version:** [bmeyer2025/glm5-reasoning-traces-sft](https://huggingface.co/datasets/bmeyer2025/glm5-reasoning-traces-sft) — same data pre-formatted with `<think>`/`<answer>` tags, split 80/10/10, ready to plug into any trainer
- **Distilled model:** [bmeyer2025/qwen3.5-4b-glm5-reasoning-distilled](https://huggingface.co/bmeyer2025/qwen3.5-4b-glm5-reasoning-distilled)
- **Code:** [brianmeyer/distillreasoning](https://github.com/brianmeyer/distillreasoning)

## License

MIT