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

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 29, 2026Architecture:Transformer Featherless Exclusive Cold

agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-base is a 4 billion parameter Qwen3-4B-Instruct-based model, fine-tuned using OpenRLHF GRPO with a 32768 token context length. This model is specifically optimized for code generation tasks, achieving a pass@8 score of 21.3434 on the cobalt-train frontier, indicating its proficiency in solving programming problems. It was developed by agurung and excels at generating correct code based on binary code-correctness reward signals.

Loading preview...

Model Overview

This model, cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-base, is an OpenRLHF GRPO (Group-normalized advantages, no KL penalty) reinforcement-learning checkpoint based on the Qwen3-4B-Instruct-2507 model. It was directly fine-tuned from the base Qwen3-4B without an initial Supervised Fine-Tuning (SFT) seed, applying RL directly to enhance its capabilities.

Key Capabilities & Optimization

  • Code Generation Excellence: This checkpoint is noted as the best by pass@8 in its development run, achieving 21.3434 on held-out validation problems. It was trained and validated on the cobalt-train ≤2/64 frontier, focusing on problems the base model struggled with.
  • Reward Signal: Training utilized a binary code-correctness reward, meaning generated programs received a 1.0 if they passed problem tests and 0.0 otherwise, directly optimizing for functional code.
  • RL Algorithm: Employs the GRPO algorithm with specific penalties for truncated samples (-1.0) and overlong responses (-0.25 ramping penalty for the last 1024 tokens before the cap).

Performance Metrics (Held-out validation, 8 samples/problem)

  • pass@1: 16.7068 (mean per-problem fraction correct)
  • pass@8: 21.3434 (problem counted solved if any sample is correct)

When to Use This Model

This model is particularly well-suited for applications requiring robust code generation, especially in scenarios where functional correctness is paramount. Its training methodology and evaluation metrics indicate strong performance in solving programming challenges.