agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot30v11v

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 30, 2026Architecture:Transformer Featherless Exclusive Cold

The agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot30v11v model is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint based on Qwen3-4B-Instruct-2507, with a context length of 32768 tokens. This model is specifically fine-tuned for code generation tasks, excelling at solving programming problems by optimizing for binary code-correctness. It achieved a pass@8 score of 5.5100 on a held-out validation set of challenging coding problems, making it suitable for applications requiring robust code generation capabilities.

Loading preview...

Model Overview

This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-groot30v11v, is a 4 billion parameter reinforcement-learning (RL) checkpoint derived from the Qwen/Qwen3-4B-Instruct-2507 base model. It was developed using the OpenRLHF GRPO algorithm, applying RL directly to the base model without an initial Supervised Fine-Tuning (SFT) seed. The model is specifically optimized for code generation, with its training focused on achieving binary code-correctness.

Key Capabilities

  • Code Generation: Excels at generating correct code, as evidenced by its training objective of binary code-correctness.
  • Reinforcement Learning Optimization: Utilizes the GRPO algorithm with specific penalties for truncated and overlong responses, enhancing generation quality.
  • Performance on Coding Benchmarks: Achieved a pass@8 score of 5.5100 and a pass@1 score of 2.5380 on a held-out validation set of 112 challenging coding problems (the "cobalt-train \u22642/64 frontier").

Good For

  • Automated Code Problem Solving: Ideal for tasks requiring the generation of functional code that passes specific tests.
  • Code Assistants: Can be integrated into tools that help developers write or debug code by suggesting correct implementations.
  • Research in RL for Code Generation: Provides a strong baseline for further experimentation in applying reinforcement learning to improve code synthesis.

Finetunes
1 models