allenai/qwen35-9b-swesmith

Overview

This model, allenai/qwen35-9b-swesmith, is a 9 billion parameter language model developed by Ai2. It is built upon the Qwen 3.5 9B base model and has been specifically fine-tuned using DPPO (Distributed Proximal Policy Optimization) to function as a terminal agent. The training focused on the SWE-smith dataset, indicating its specialization in automated software engineering tasks.

Key Capabilities

• Terminal Agent Functionality: Designed to operate effectively within a terminal environment, enabling automated task execution.
• Software Engineering Focus: Trained on the SWE-smith dataset, suggesting proficiency in understanding and generating code-related solutions.
• DPPO Fine-tuning: Utilizes advanced reinforcement learning techniques for improved performance in interactive, agent-based scenarios.
• Part of Tmax Collection: This model is one of several terminal agents developed by AllenAI, offering a range of specialized models for different terminal-based applications.

Performance Highlights

Evaluation results show its performance on terminal-based benchmarks:

• TB Lite: Achieved 47.2 ± 2.2 on the TB Lite benchmark.
• TB 2.1: Scored 21.0 ± 0.5 on the TB 2.1 benchmark.

These metrics position it as a capable agent for specific terminal interaction tasks, though other models in the Tmax collection, such as Tmax 9B, show higher overall scores on these benchmarks.

Training Details

• Base Model: Fine-tuned from hamishivi/Qwen3.5-9B.
• Dataset: Utilized the SWE-smith dataset.
• Hyperparameters: Trained with a max prompt token limit of 2048, max per-turn tokens of 16384, and a total of 500 training steps, with the best checkpoint identified at 200 steps based on TBLite performance.

Good For

• Developing and experimenting with terminal agents.
• Automating tasks within software development workflows.
• Research into reinforcement learning for language models in interactive environments.