agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30
agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30 is a 4 billion parameter Qwen3-based causal language model, fine-tuned using OpenRLHF GRPO for enhanced code generation. This model excels at solving programming problems, achieving a pass@8 score of 1.0587 on held-out validation problems. It is specifically optimized for code correctness, making it suitable for tasks requiring reliable program output.
Loading preview...
Overview
This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-iid30, is a 4 billion parameter Qwen3-based language model that has undergone Reinforcement Learning (RL) fine-tuning using the OpenRLHF GRPO algorithm. It is built upon the Qwen/Qwen3-4B-Instruct-2507 base model, with RL applied directly to the base model without an initial Supervised Fine-Tuning (SFT) seed. This checkpoint represents the best performing iteration in its training run, specifically optimized for code generation tasks.
Key Capabilities
- Enhanced Code Generation: Fine-tuned to produce correct code, validated by a binary code-correctness reward signal (1.0 for passing tests, 0.0 otherwise).
- Strong Performance on Code Problems: Achieved a pass@8 score of 1.0587 and a pass@1 score of 0.4102 on held-out validation problems, indicating its proficiency in solving programming challenges.
- RL-Optimized: Utilizes the GRPO algorithm with group-normalized advantages and specific penalties for truncated or overlong responses, ensuring robust and efficient learning.
Good For
- Code Generation: Ideal for applications requiring the generation of functional and correct code snippets or programs.
- Automated Problem Solving: Suitable for tasks involving solving programming problems, especially those where correctness can be verified through tests.
- Research in RL for Code: Provides a strong baseline or comparison point for further research into applying reinforcement learning to code generation models.