FIM-Mid-14B is the mid-trained checkpoint of the FIM 14B pipeline: Qwen2.5-Coder-14B-Instruct after function-aware FIM mid-training, before agent post-training. Post-training this checkpoint on R2E-Gym agent trajectories produces TIGER-Lab/FIM-14B.

It is released for reproducibility and further post-training. The paper deliberately never scores mid-training-only checkpoints — a FIM-only model has degraded instruction-following and cannot be compared fairly against instruction-tuned baselines; every reported gain is one that survives post-training.

Training

Base model: Qwen/Qwen2.5-Coder-14B-Instruct
FIM mid-training: midtraining/configs/fim_midtrain.yaml on TIGER-Lab/FIM-Midtraining-400K — AdamW, lr 1.0e-5, cosine schedule, warmup ratio 0.1, weight decay 0.05, one epoch, sequence length 32768, bf16 (as-run copy: FIM_Midtrain_14B.yaml)
Post-training: none — see TIGER-Lab/FIM-14B for the post-trained agent model

Serve with vLLM

A standard Qwen2.5 checkpoint; no overrides needed at its native 32768 context:

bash

CUDA_VISIBLE_DEVICES=0 \
python -m vllm.entrypoints.openai.api_server \
  --model TIGER-Lab/FIM-Mid-14B \
  --served-model-name FIM-Mid-14B \
  --host 127.0.0.1 \
  --port 8400 \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.9 \
  > vllm_fim_mid14b.log 2>&1 &

Post-training

To reproduce FIM-14B, run R2E-Gym trajectory SFT from this checkpoint — the exact config is posttraining/r2egym/FIM_Posttrain_14B.yaml (LLaMA-Factory, full fine-tuning, lr 1.0e-5, 2 epochs, cutoff 32768), which already points at this repo id. See posttraining/r2egym/ for the walkthrough.

Citation

bibtex

@article{wang2026fim,
  title={Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models},
  author={Wang, Yubo and Liang, Jiarong and Zhang, Yuxuan and Liu, Xuye and Wei, Cong and Zhang, Yuyu and Nie, Ping and Chen, Wenhu},
  journal={arXiv preprint arXiv:2607.12463},
  year={2026}
}