FineEnvs/Qwen3.5-2B-opencode-RL
FineEnvs/Qwen3.5-2B-opencode-RL is a 2.3 billion parameter language model, fine-tuned from Qwen/Qwen3.5-2B by FineEnvs using asynchronous GRPO (TRL Async GRPO) with OpenCode. This model is specifically optimized for agentic data-analysis tasks, demonstrating a 39.5% pass@1 across four evaluation harnesses. Its primary strength lies in its ability to perform text/tool interactions for data analysis, making it suitable for applications requiring automated data processing and agentic reasoning.
Loading preview...
Qwen3.5-2B-opencode-RL: Agentic Data Analysis Model
FineEnvs/Qwen3.5-2B-opencode-RL is a 2.3 billion parameter model derived from Qwen/Qwen3.5-2B, specifically fine-tuned for agentic data-analysis tasks. It leverages asynchronous GRPO (TRL Async GRPO) with OpenCode for its training methodology. This model achieved a 39.5% pass@1 across four evaluation harnesses (OpenCode, Claude Code, Codex, Mini-SWE-Agent) on 250 fixed SmolDataEnvs test tasks.
Key Capabilities
- Agentic Data Analysis: Optimized for tasks requiring interaction with tools and environments for data processing.
- Reinforcement Learning Fine-tuning: Utilizes TRL Async GRPO with binary correctness-only rewards, initialized from the base Qwen model.
- OpenCode Integration: Training rollouts exclusively use OpenCode, running through OpenEnv \u00d7 Harbor in E2B sandboxes.
- Performance: Demonstrated a 40.0% pass@1 on the OpenCode harness specifically.
Good For
- Automated Data Processing: Ideal for applications that require models to analyze and interact with data programmatically.
- Agent Development: Suitable for building and evaluating AI agents focused on data-centric problem-solving.
- Research in RL for Code Generation: Provides a strong baseline for exploring reinforcement learning techniques in code-related tasks, particularly with OpenCode environments.