TIGER-Lab/FIM-7B

FIM-7B: Enhanced Coding Agent Model

FIM-7B is a 7.6 billion parameter coding-agent model from TIGER-Lab, specifically designed to excel in automated code generation and repair tasks. It is built upon the Qwen2.5-Coder-7B-Instruct base model and distinguishes itself through a unique "mid-training" stage.

Key Capabilities & Training

• Function-Aware Fill-in-the-Middle (FIM) Mid-Training: The core innovation of FIM-7B is its mid-training on a specialized dataset, TIGER-Lab/FIM-Midtraining-400K, which incorporates function-aware FIM data. This process optimizes the model for understanding and completing code segments within a larger functional context.
• Post-Training with R2E-Gym Trajectories: Following mid-training, the model undergoes post-training using R2E-Gym agent trajectories, further refining its agentic capabilities for problem-solving.
• Significant Performance Gains: This specialized training methodology yields substantial improvements on critical coding benchmarks:
  • +2.8 points on SWE-Bench-Verified.
  • +3.7 points on SWE-Bench-Lite.

When to Use FIM-7B

• Automated Code Repair: Its strong performance on SWE-Bench benchmarks indicates high proficiency in identifying and fixing bugs in existing codebases.
• Code Generation and Completion: The FIM mid-training makes it particularly adept at generating contextually relevant code snippets and completing partial code.
• Coding Agent Development: Ideal for integration into coding agent scaffolds, especially those requiring robust problem-solving and code manipulation capabilities.