agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot16
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot16 is a 4 billion parameter Qwen3-4B model fine-tuned using OpenRLHF GRPO for improved code generation. This reinforcement learning checkpoint, derived from Qwen/Qwen3-4B-Instruct-2507, is specifically optimized for solving programming problems. It was trained on a frontier of challenging code problems, using binary code-correctness as the reward signal, making it suitable for tasks requiring accurate code output.
Loading preview...
Model Overview
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot16 is a 4 billion parameter language model based on the Qwen/Qwen3-4B-Instruct-2507 architecture. It is a reinforcement learning (RL) checkpoint, specifically fine-tuned using the OpenRLHF GRPO algorithm. This model was seeded directly from the base Qwen3-4B model, without an initial Supervised Fine-Tuning (SFT) step, and represents the best performing checkpoint by pass@8 in its training run.
Key Capabilities
- Code Generation: Optimized for generating correct code, with training focused on solving programming problems.
- Reinforcement Learning: Utilizes GRPO (group-normalized advantages, no KL penalty) for fine-tuning, enhancing performance on specific tasks.
- Code-Correctness Reward: Trained with a binary code-correctness reward signal, meaning it's rewarded for generating programs that pass problem tests.
- Anti-Truncation Shaping: Incorporates a stop-properly penalty for truncated samples and an overlong penalty for responses nearing the token cap, aiming for complete and concise outputs.
Use Cases
This model is particularly well-suited for:
- Automated Code Problem Solving: Generating solutions for programming challenges.
- Code Completion and Generation: Tasks where the correctness of the generated code is paramount.
- Research in RL for Code: As a checkpoint demonstrating GRPO application for code generation.