agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-base-b20
The agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-base-b20 model is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint for the Qwen3-4B architecture, fine-tuned for code generation. It was seeded directly from the base Qwen3-4B model and trained using a binary code-correctness reward signal. This model excels at solving programming problems, particularly those identified as challenging for the base model, and is optimized for generating correct code outputs within a 32768 token context window.
Loading preview...
Model Overview
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-base-b20 is a 4 billion parameter model derived from the Qwen3-4B architecture, specifically an OpenRLHF GRPO (Group-Normalized Advantages, No KL Penalty) reinforcement-learning checkpoint. Unlike many models, it was seeded directly from the base Qwen3-4B model without an initial supervised fine-tuning (SFT) phase, applying RL directly.
Key Capabilities
- Code Generation: This model is highly specialized for generating correct code, having been trained with a binary code-correctness reward signal (1.0 for passing tests, 0.0 otherwise).
- Problem Solving: It was trained and validated on a challenging subset of problems (the "cobalt-train \u22642/64 frontier") where the base model struggled, indicating its proficiency in tackling difficult coding tasks.
- RL Optimization: Utilizes the GRPO algorithm with specific penalties for truncated samples and overlong responses, aiming for robust and complete code outputs.
Good For
- Program Synthesis: Ideal for tasks requiring the generation of functional code snippets or solutions to programming problems.
- Improving Base Model Performance: Particularly effective for scenarios where the base Qwen3-4B model shows limitations in code correctness.
- Research in RL for Code: Provides a checkpoint for studying the effects of GRPO and specific reward shaping techniques on code generation.