agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bgroot8
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bgroot8 is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint for the Qwen3-4B base model, developed by agurung. This model is specifically optimized for code generation tasks, particularly for solving programming problems by achieving binary code-correctness. It was trained on the cobalt-train frontier dataset, focusing on problems the base model struggled with, and is noted as the best checkpoint by pass@8 in its development run.
Loading preview...
Model Overview
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bgroot8 is a 4 billion parameter reinforcement-learning (RL) checkpoint derived from the Qwen/Qwen3-4B-Instruct-2507 base model. Developed by agurung, this model utilizes the OpenRLHF GRPO algorithm and was seeded directly from the base Qwen3-4B without an initial Supervised Fine-Tuning (SFT) step. It was saved at global step 60 of its RL training run and is identified as the best performing checkpoint by pass@8 within that run.
Key Capabilities & Training
This model is primarily designed for code generation, with its training focused on achieving binary code-correctness. The reward signal during training was 1.0 for generated programs that passed problem tests and 0.0 otherwise. It was trained and validated on the cobalt-train ≤2/64 frontier dataset, which comprises 1833 training and 112 held-out validation problems that the base model solved on at most 2 of 64 samples. The training incorporated specific techniques like a stop-properly penalty for truncated samples (-1.0 reward) and a DAPO overlong penalty for responses nearing the token cap.
When to Use This Model
This model is particularly well-suited for use cases requiring high-accuracy code generation, especially for solving programming challenges where a correct, executable solution is the primary objective. Its optimization for pass@8 on challenging problems suggests strong performance in generating functional code. Developers can load it directly using AutoModelForCausalLM and AutoTokenizer from the transformers library or serve it with vLLM.