allenai/qwen35-9b-cli-gym
The allenai/qwen35-9b-cli-gym is a 9 billion parameter language model developed by Ai2, fine-tuned from Qwen 3.5 9B using DPPO. It is specifically optimized as a terminal-agent for command-line interface (CLI) tasks, trained on the CLI-Gym dataset. This model is designed for use in environments requiring automated interaction with terminal interfaces, demonstrating performance on benchmarks like TB Lite and TB 2.1.
Loading preview...
Overview
This model, allenai/qwen35-9b-cli-gym, is a 9 billion parameter terminal-agent developed by Ai2. It is fine-tuned from the Qwen 3.5 9B base model using the DPPO (Distributed Proximal Policy Optimization) method. The primary goal of this model is to function effectively as an agent within command-line interface (CLI) environments, having been specifically trained on the CLI-Gym dataset.
Key Capabilities
- Terminal Agent Functionality: Optimized for interacting with and performing tasks within terminal environments.
- DPPO Fine-tuning: Utilizes DPPO for enhanced performance in agentic tasks.
- CLI-Gym Dataset: Trained on a specialized dataset for CLI interactions, ensuring relevance to terminal-based use cases.
- Evaluation Performance: Achieves competitive results on terminal-agent benchmarks such as TB Lite (50.7 ± 5.9) and TB 2.1 (25.1 ± 1.4).
Training Details
The model was trained with specific hyperparameters including a maximum prompt token length of 2048, a maximum per-turn token length of 16384, and a total training of 500 steps, with the best performing checkpoint at 200 steps on TBLite. It uses a constant learning rate of 1e-6 and an Apache 2.0 license.
Good For
- Developing and testing automated terminal agents.
- Research into reinforcement learning for interactive environments.
- Applications requiring a language model to interpret and execute command-line instructions.