33 lines
1.5 KiB
Markdown
33 lines
1.5 KiB
Markdown
|
---
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen3-0.6B
|
||
|
|
datasets:
|
||
|
|
- taki555/DeepScaleR-Easy
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
license: apache-2.0
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
---
|
||
|
|
|
||
|
|
# Qwen3-0.6B-Art
|
||
|
|
|
||
|
|
This is the Chain-of-Thought (CoT) efficient version of the [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) model, trained on the [DeepScaleR-Easy](https://huggingface.co/datasets/taki555/DeepScaleR-Easy) dataset.
|
||
|
|
|
||
|
|
This model was introduced in the paper [The Art of Efficient Reasoning: Data, Reward, and Optimization](https://huggingface.co/papers/2602.20945). Check the [Project Page](https://wutaiqiang.github.io/project/Art) for more details.
|
||
|
|
|
||
|
|
## Model Description
|
||
|
|
Large Language Models (LLMs) consistently benefit from scaled Chain-of-Thought (CoT) reasoning but also suffer from heavy computational overhead. Efficient reasoning aims to incentivize short yet accurate thinking trajectories, typically through reward shaping with Reinforcement Learning (RL).
|
||
|
|
|
||
|
|
This model follows a two-stage training paradigm: length adaptation and reasoning refinement. It is optimized to maintain a sufficient density of positive reward signals while avoiding the "short-is-correct" trap, demonstrating robust and generalized efficient reasoning capabilities.
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
```bibtex
|
||
|
|
@inproceedings{wu2026art,
|
||
|
|
title={The Art of Efficient Reasoning: Data, Reward, and Optimization},
|
||
|
|
author={Taiqiang Wu and Zenan Xu and Bo Zhou and Ngai Wong},
|
||
|
|
year={2026},
|
||
|
|
url={https://arxiv.org/pdf/2602.20945}
|
||
|
|
}
|
||
|
|
```
|