wizardII/ArcherCodeR-1.5B-DAPO

✨ ArcherCodeR

🏹️ Reinforcement Learning for Enhanced Code Reasoning in LLMs 🎯

Overview

_{ArcherCodeR-1.5B-DAPO achieves progressive improvements on LiveCodeBench (LCB), reaching 27.24% LCB score.}

ArcherCodeR is an open-source initiative enhancing code reasoning in large language models through scalable, rule-governed reinforcement learning. We provide full-stack reproducibility including:

• Training code and pipelines
• Curated datasets
• Trained models
• Complete training logs

Current Models:

• ArcherCodeR-1.5B-DAPO - achieves state-of-the-art performance on code tasks (LiveCodeBench) among comparable-scale models (excluding our final ArcherCodeR-1.5B). All training components for this model are now fully released.
• ArcherCodeR-1.5B - SOTA among similarly-sized models (training pipeline releasing progressively)

Evaluation

Performance on LiveCodeBench. The Pass@1 metric represents the average performance across 4 independent sampling attempts. To ensure consistency, we re-evaluated all comparable open-source models using identical evaluation scripts and parameters (temperature=0.8, max_gen_length=32k).

The detailed results are shown in the table below.

Model	LCB (8/1/24-2/1/25)(Pass@1)	LCB (8/1/24-2/1/25)(Pass@4)
DeepSeek-R1-Distill-Qwen-1.5B	16.9	—
DeepSeek-R1-Distill-Qwen-1.5B(Tested)	16.40	25.81
DeepCoder-1.5B	25.1	—
DeepCoder-1.5B(Tested)	23.03	30.82
Nemotron-Research-Reasoning-Qwen-1.5B	23.81	—
Nemotron-Research-Reasoning-Qwen-1.5B(Tested)	25.45	34.40
ArcherCodeR-1.5B-DAPO	26.70	36.56
ArcherCodeR-1.5B(32k)	28.49	38.71
ArcherCodeR-1.5B(48k)	29.30	39.07

Note:

1. Evaluation variance for the same model is typically within ±0.5 across multiple runs.
2. DeepCoder consistently scored around 23 in our tests - lower than its reported performance.
3. NVIDIA's Nemotron-Research-Reasoning-Qwen-1.5B slightly outperformed its reported score, potentially due to different parameter settings in their original evaluation.

Technical Report

Coming soon.

Acknowledgements

• We build our model upon DeepSeek-R1-Distill-Qwen-1.5B.
• Training was carried out with a modified version of verl.