agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp20-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

agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp20-groot30v11v is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint for the Qwen3-4B base model, developed by agurung. This model is specifically fine-tuned for code generation, optimizing for binary code-correctness. It excels at solving programming problems, particularly those at the cobalt-train frontier, making it suitable for code-centric applications requiring high accuracy.

Loading preview...

Model Overview

This model, developed by agurung, is an OpenRLHF GRPO reinforcement-learning checkpoint based on the Qwen/Qwen3-4B-Instruct-2507 architecture. It is a 4 billion parameter model with a 32768 token context length, directly applying RL to the base Qwen3-4B model without an initial Supervised Fine-Tuning (SFT) seed.

Key Capabilities

  • Code Generation Optimization: The model is specifically trained and validated on the cobalt-train ≤2/64 frontier, focusing on programming problems that the base model solved on at most 2 of 64 samples.
  • Reward Signal: Training utilizes a binary code-correctness reward signal, assigning 1.0 for passing problem tests and 0.0 otherwise, directly optimizing for functional code.
  • Performance: This checkpoint is identified as the best by pass@8 in its training run, indicating strong performance in generating correct code solutions.
  • RL Algorithm: Employs the GRPO algorithm, which includes group-normalized advantages and no KL penalty, alongside specific anti-truncation and overlong penalties.

Use Cases

This model is particularly well-suited for tasks requiring robust and accurate code generation, especially for competitive programming challenges or automated code problem-solving. Its training methodology and reward signal make it a strong candidate for applications where the correctness of generated code is paramount.