Dataset · mithulaartigala
Millions of reasoning chains from various LLMs
mithulaartigala/reasoning-corpus-4K-5M-v1
Filtered reasoning traces from DeepSeek, Qwen, Gemma, and other models for small model distillation.
What's inside
A corpus of thought traces from several models like DeepSeek-R1 variants, Qwen3, and others, gathered from multiple repositories and filtered for training small language models. The language is English, volume ranges from 1 to 10 million examples, and length is within 5k tokens. The card lists repo_id, tok_len, user, thought_trace, assistant, and ChatML columns.
What you can do with it
- Distill reasoning into a small model via SFT
- Train LoRA on reasoning chains
- Filter and analyze traces by length and source
- Tasks
- Text generation
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 6.1K
- Likes
- 213
How to load it
from datasets import load_dataset
ds = load_dataset("mithulaartigala/reasoning-corpus-4K-5M-v1")