allenai/tmax-sft-8b
allenai/tmax-sft-8b is an 8 billion parameter instruction-tuned language model developed by Ai2, fine-tuned from Qwen 3 8B. This model is specifically designed and optimized for use as a terminal agent, demonstrating strong performance on terminal-based tasks. It serves as the supervised fine-tuning base for the Tmax 8B RL-trained model, excelling in environments requiring command-line interaction.
Loading preview...
Overview
allenai/tmax-sft-8b is an 8 billion parameter language model developed by Ai2, specifically designed for use as a terminal agent. It is built upon the Qwen 3 8B architecture and has undergone supervised fine-tuning (SFT) using the TMax-15k dataset. This model is a foundational component for the subsequent RL training of the Tmax 8B model, focusing on enhancing its capabilities in terminal environments.
Key Capabilities & Performance
- Terminal Agent Optimization: Tmax SFT 8B is explicitly trained to function as a terminal agent, enabling it to interact effectively with command-line interfaces.
- Evaluation Benchmarks: It demonstrates improved performance over its base model, Qwen 3 8B, on terminal-based evaluation tasks. Specifically, it achieves 11.5 +/- 0.1 on TB Lite and 6.0 +/- 1.4 on TB 2.1, outperforming Qwen 3 8B's scores of 7.3 +/- 1.0 and 1.1 +/- 0.9 respectively.
- Training Details: The model was trained with a maximum overall token length of 32768, a global train batch size of 128, and 2 epochs, using a linear cooldown learning rate scheduler with 0.03% warmup.
When to Use This Model
- Terminal Automation: Ideal for applications requiring an AI to interact with and execute commands within a terminal environment.
- Base for Further Training: Suitable as a strong SFT base model for researchers and developers looking to conduct further reinforcement learning or specialized fine-tuning for terminal agent tasks.
- Research in Agentic LLMs: Valuable for academic and research purposes exploring the development and evaluation of language models as autonomous agents.