agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bvs8

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-bvs8 is a 4 billion parameter OpenRLHF GRPO reinforcement-learning checkpoint for Qwen3-4B, specifically fine-tuned for code generation. This model was seeded from the base Qwen3-4B and optimized using a binary code-correctness reward signal. It excels at generating correct code, having been identified as the best checkpoint by pass@8 in its training run, making it suitable for programming assistance and automated code generation tasks.

Loading preview...

Model Overview

The agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bvs8 is a 4 billion parameter reinforcement-learning checkpoint based on the Qwen/Qwen3-4B-Instruct-2507 model. Developed by agurung, this model utilizes the OpenRLHF GRPO algorithm and was trained without an SFT seed, applying RL directly to the base Qwen3-4B.

Key Capabilities

  • Optimized for Code Correctness: The model's primary strength lies in generating functionally correct code, as evidenced by its "best by pass@8" metric during training.
  • Reinforcement Learning: Trained using a GRPO algorithm with a binary code-correctness reward signal, meaning it learns directly from whether its generated code passes tests.
  • Anti-Truncation Shaping: Incorporates a "stop-properly penalty" where truncated samples receive a -1.0 reward, and an "overlong penalty" for responses nearing the maximum token limit, encouraging complete and concise outputs.

Good For

  • Code Generation: Ideal for tasks requiring the generation of executable and correct code, particularly for problems similar to those in the cobalt-train frontier dataset.
  • Programming Assistance: Can be used in applications where a model needs to produce high-quality, test-passing code snippets or solutions.

This checkpoint represents a specific stage (global step 48) of an RL training run, focusing on maximizing code pass rates.