agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30v11v
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30v11v is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint for Qwen3-4B, developed by agurung. This model is specifically fine-tuned for code generation, excelling at producing correct code solutions as measured by binary code-correctness. It demonstrates strong performance on the cobalt-train frontier, achieving a pass@8 score of 7.4901 on held-out validation problems.
Loading preview...
Model Overview
This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30v11v, is a 4 billion parameter reinforcement-learning checkpoint based on the Qwen/Qwen3-4B-Instruct-2507 architecture. It was developed by agurung using the OpenRLHF GRPO algorithm, with RL applied directly to the base Qwen3-4B model without an intermediate SFT seed.
Key Capabilities & Training
- Code Generation: The model is specifically optimized for generating correct code, using a binary code-correctness reward signal (1.0 for passing tests, 0.0 otherwise).
- Performance: It achieved a pass@8 score of 7.4901 and a pass@1 score of 4.2843 on held-out validation problems, making it the best checkpoint by pass@8 in its development run.
- Training Data: Trained and validated on the cobalt-train \u22642/64 frontier, comprising 1833 training and 112 held-out validation problems that the base model solved on at most 2 of 64 samples.
- RL Recipe: Utilizes the GRPO algorithm with group-normalized advantages and no KL penalty. It incorporates a stop-properly penalty (-1.0 for truncated samples) and an overlong penalty for responses in the last 1024 tokens before the cap.
Use Cases
This model is particularly well-suited for tasks requiring high-quality code generation and problem-solving, especially in competitive programming or automated code assessment scenarios where correctness is paramount. Its fine-tuning on a specific code-centric dataset makes it a strong candidate for applications demanding robust and accurate code outputs.