agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-miid16

Model Overview

agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-miid16 is a 4 billion parameter Qwen3-4B model that has undergone Reinforcement Learning (RL) using the OpenRLHF GRPO algorithm. It is based on Qwen/Qwen3-4B-Instruct-2507 but was seeded directly from the base Qwen3-4B model without an initial Supervised Fine-Tuning (SFT) step.

Key Capabilities

• Code Generation: Optimized for generating correct code, as evidenced by its training on a binary code-correctness reward signal (1.0 for passing tests, 0.0 otherwise).
• Problem Solving: Specifically trained and validated on the "cobalt-train \u22642/64 frontier," focusing on problems that the base model could solve on at most 2 of 64 samples, indicating an improvement in tackling challenging coding tasks.
• RL-Enhanced Performance: This checkpoint represents the "best by pass@8" in its training run, suggesting strong performance in generating functional code.

Training Details

The model was trained using the GRPO algorithm, which incorporates group-normalized advantages and no KL penalty. It includes a "stop-properly penalty" where truncated samples receive a -1.0 reward and a DAPO overlong penalty for responses nearing the token limit. The training involved 2 episodes with 8 samples per prompt and a max of 4096 new tokens per rollout.

Good For

• Automated Code Generation: Ideal for applications requiring the generation of functional and correct code snippets.
• Programming Assistance: Can be used in tools for code completion, bug fixing, or generating solutions to defined programming problems.