agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp5-q4v3-groot16
The agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp5-q4v3-groot16 is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint based on Qwen3-4B-Instruct-2507, with a context length of 32768 tokens. This model is specifically fine-tuned for code generation, utilizing a binary code-correctness reward signal. It excels at solving programming problems, demonstrating strong performance on the cobalt-train frontier with a pass@8 score of 1.8061.
Loading preview...
Overview
This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp5-q4v3-groot16, is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint derived from the Qwen/Qwen3-4B-Instruct-2507 base model. It was developed by agurung and represents the best performing checkpoint (by pass@8) from its training run, saved at global step 24.
Key Capabilities
- Code Generation: Specifically optimized for generating correct code, trained with a binary code-correctness reward signal.
- Problem Solving: Demonstrates strong performance on programming challenges, validated against the
cobalt-trainfrontier. - Reinforcement Learning: Utilizes the GRPO algorithm with specific penalties for truncated and overlong responses to improve output quality.
Performance Metrics
Evaluated on a held-out validation set with 8 samples per problem, this checkpoint achieved:
- pass@1:
0.2794(unbiased mean per-problem fraction correct) - pass@8:
1.8061(problem counted solved if any sample is correct)
Good for
- Automated Code Generation: Ideal for tasks requiring the model to produce functional code that passes tests.
- Competitive Programming Assistance: Can be used to generate solutions for programming problems, particularly those on the
cobalt-trainfrontier. - Research in RL for Code: Provides a strong baseline for further research into reinforcement learning applications for code generation.