agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-biid16

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

The agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-biid16 model is a 4 billion parameter Qwen3-4B-Instruct-based reinforcement learning checkpoint, specifically optimized for code generation tasks. Developed by agurung, this model was trained using the OpenRLHF GRPO algorithm with a binary code-correctness reward signal. It excels at solving programming problems, particularly those on the cobalt-train frontier, demonstrating strong performance in generating correct code.

Loading preview...

Model Overview

This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-biid16, is a 4 billion parameter reinforcement learning (RL) checkpoint based on the Qwen/Qwen3-4B-Instruct-2507 model. It was developed by agurung using the OpenRLHF GRPO algorithm, with RL applied directly to the base Qwen3-4B model without an initial Supervised Fine-Tuning (SFT) seed.

Key Capabilities & Training

  • Code Generation Optimization: The model is specifically optimized for generating correct code, with training focused on a binary code-correctness reward signal (1.0 for passing tests, 0.0 otherwise).
  • Performance: It is identified as the best checkpoint by pass@8 in its training run, indicating strong performance in generating functional code solutions.
  • Targeted Training Data: Training and validation were conducted on the "cobalt-train \u22642/64 frontier," comprising 1833 training and 112 held-out validation problems that the base model solved on at most 2 of 64 samples.
  • Advanced RL Techniques: The training utilized GRPO (group-normalized advantages, no KL penalty), incorporating a stop-properly penalty (-1.0 for truncated samples) and an overlong penalty for responses nearing the token cap.

Use Cases

This model is particularly well-suited for:

  • Code Generation: Generating functional code snippets or solutions to programming problems.
  • Automated Programming Assistance: Tasks requiring high accuracy in code output, especially for competitive programming or problem-solving scenarios.
  • Research in RL for Code: As a strong example of an RL-tuned model for code correctness, it can be a valuable baseline for further research.

Finetunes
1 models