agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bvs16

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

agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bvs16 is a 4 billion parameter Qwen3-based model, fine-tuned using OpenRLHF GRPO. This model was specifically optimized for code generation tasks, applying reinforcement learning directly to the base model with a binary code-correctness reward signal. It is designed to produce correct code solutions, having been trained and validated on a challenging set of programming problems. The model's primary strength lies in its ability to generate functional code, achieving the best pass@8 performance in its development run.

Loading preview...

Model Overview

This model, agurung/cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp10-bvs16, is a 4 billion parameter Qwen3-based language model. It has been fine-tuned using the OpenRLHF GRPO (Group-Normalized Advantages, No KL Penalty) reinforcement learning algorithm. Notably, RL was applied directly to the base Qwen/Qwen3-4B-Instruct-2507 model without an initial Supervised Fine-Tuning (SFT) seed.

Key Capabilities

  • Code Generation: The model is specifically optimized for generating correct code. Its training utilized a binary code-correctness reward signal, meaning it was rewarded for producing programs that pass problem-specific tests.
  • Performance on Code Problems: This checkpoint achieved the best pass@8 metric during its development run, indicating a strong ability to generate functional code solutions.
  • Targeted Training: It was trained and validated on the "cobalt-train \u22642/64 frontier," a dataset of challenging programming problems that the base model struggled with.

Training Details

The model was trained for 2 episodes with a maximum of 4096 new tokens per rollout. It incorporates a stop-properly penalty (-1.0 for truncated samples) and a DAPO overlong penalty ramping to -0.25 for responses in the last 1024 tokens before the cap. The actor learning rate was constant at 1e-06.

Good for

  • Code generation tasks where functional correctness is paramount.
  • Developers looking for a Qwen3-4B variant specifically enhanced for programming challenges through reinforcement learning.

Finetunes
1 models