laion/Qwen3-32B-R2EGYM-256-3epochs
laion/Qwen3-32B-R2EGYM-256-3epochs is a 32 billion parameter Qwen3-based causal language model, fine-tuned by laion using the SkyRL framework. This model specializes in coding and reasoning tasks, leveraging asynchronous PPO on the R2EGYM benchmark. It is optimized for terminal-based coding agent interactions within sandboxed environments, offering enhanced performance for programming challenges.
Loading preview...
Model Overview
This model, laion/Qwen3-32B-R2EGYM-256-3epochs, is a reinforcement learning fine-tuned variant of the 32 billion parameter Qwen/Qwen3-32B base model. It was developed by laion using the SkyRL framework, which employs a fully asynchronous PPO strategy.
Key Capabilities & Training
- Specialized Fine-tuning: The model is specifically trained on coding and reasoning tasks derived from the R2EGYM benchmark, utilizing a subset of 256 tasks from
open-athena/r2egym_gpt5_codex_solved_tasks_256_subset. - Agent-Based Interaction: It integrates with Terminus-2, a terminal-based coding agent, enabling it to interact with sandboxed Docker environments to solve programming problems. Training included an enabled "thinking mode" for this agent.
- Reinforcement Learning: The fine-tuning process involved 3 epochs and 12 total steps, using FSDP2 with CPU offload for parallelism.
- Extended Context: The model supports a maximum generate length of 30,720 tokens and utilizes YARN RoPE scaling with a factor of 4.0, building upon the original 32,768 token context of Qwen3-32B.
Use Cases
This model is designed as a direct replacement for Qwen3-32B, offering improved capabilities in:
- Coding task resolution
- Automated reasoning in programming contexts
- Interacting with terminal-based coding environments