TIGER-Lab/FIM-Mid-8B
FIM-Mid-8B is an 8 billion parameter language model developed by TIGER-Lab, serving as a mid-trained checkpoint of the FIM 8B pipeline. It is based on Qwen3-8B and has undergone function-aware Fill-in-the-Middle (FIM) mid-training, designed for further post-training into agent models. This model is released for reproducibility and as a foundation for developing advanced coding agents, supporting a native context length of 40960 tokens.
Loading preview...
FIM-Mid-8B: A Foundation for Coding Agents
FIM-Mid-8B is an 8 billion parameter model from TIGER-Lab, representing a crucial mid-trained checkpoint in the development of advanced coding agent models. It is derived from the Qwen3-8B base model and has undergone specialized function-aware Fill-in-the-Middle (FIM) mid-training.
Key Characteristics
- Mid-trained Checkpoint: This model is specifically released before agent post-training, making it ideal for researchers and developers looking to reproduce or build upon the FIM 8B pipeline.
- Function-Aware FIM Training: The model was trained on the TIGER-Lab/FIM-Midtraining-400K dataset with a focus on Fill-in-the-Middle capabilities, which are essential for code generation and completion tasks.
- High Context Length: It supports a native Qwen3 context length of 40960 tokens, providing ample capacity for complex coding tasks.
- Foundation for FIM-8B: Post-training this checkpoint with SWE-Lego trajectories leads to the fully developed TIGER-Lab/FIM-8B agent model.
Intended Use
FIM-Mid-8B is primarily intended for:
- Reproducibility: Allowing researchers to replicate the FIM 8B pipeline's mid-training stage.
- Further Post-training: Serving as a robust base for fine-tuning into specialized coding agent models, particularly for tasks requiring strong FIM capabilities.
- Research and Development: Exploring the impact of function-aware FIM training on model performance before instruction-following post-training.