Lite-Coder/LiteCoder-Terminal-4b-sft

LiteCoder-Terminal-4b-sft Overview

LiteCoder-Terminal-4b-sft is a specialized 4 billion parameter model from Lite-Coder, fine-tuned from Qwen3-4B-Instruct-2507. Its primary focus is on enabling lightweight code agents to operate effectively within terminal environments. The model was trained on an extensive dataset, LiteCoder-Terminal-SFT, which comprises 11,255 diverse trajectories and agent scaffolds, significantly expanding upon previous iterations.

Key Capabilities & Performance

• Terminal Agent Proficiency: Designed for agentic tasks within terminal interfaces, showing consistent improvements over its base model and previous preview versions.
• Enhanced Training Data: Benefits from a scaled-up training dataset incorporating a broader task taxonomy.
• Benchmark Improvements: Demonstrates notable performance gains across various Terminal Bench evaluations, including Terminal Bench 1.0, 2.0, and Pro. For instance, it achieves 13.44% pass@1 and 30% pass@4 on Terminal Bench 1.0, and 15.5% pass@1 on Terminal Bench Pro, outperforming the base Qwen3-4B-Instruct model significantly.

Ideal Use Cases

• Automated Terminal Operations: Suitable for developing agents that interact with and automate tasks in command-line interfaces.
• Lightweight Code Agents: Optimized for scenarios requiring efficient, smaller-scale code execution and interaction within terminal environments.
• Research in Agentic LLMs: Provides a strong baseline for further research and development in language agents operating in constrained environments.