初始化项目，由ModelHub XC社区提供模型

Model: TIGER-Lab/FIM-Mid-8B Source: Original Platform
2026-07-19 08:19:11 +08:00
commit 00b717fc0d
16 changed files with 152374 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,59 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model:
+- Qwen/Qwen3-8B
+datasets:
+- TIGER-Lab/FIM-Midtraining-400K
+tags:
+- code
+- software-engineering
+- fim
+---
+
+# FIM-Mid-8B
+
+[📄 Paper](https://arxiv.org/abs/2607.12463) · [💻 GitHub](https://github.com/TIGER-AI-Lab/FIM-Midtraining) · [🤗 Dataset](https://huggingface.co/datasets/TIGER-Lab/FIM-Midtraining-400K) · [🤗 Collection](https://huggingface.co/collections/TIGER-Lab/fim-midtraining)
+
+**FIM-Mid-8B** is the mid-trained checkpoint of the FIM 8B pipeline: `Qwen3-8B` after function-aware FIM mid-training, **before** agent post-training. Post-training this checkpoint on SWE-Lego trajectories produces [TIGER-Lab/FIM-8B](https://huggingface.co/TIGER-Lab/FIM-8B).
+
+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/Qwen3-8B`](https://huggingface.co/Qwen/Qwen3-8B)
+- **FIM mid-training**: [`midtraining/configs/fim_midtrain.yaml`](https://github.com/TIGER-AI-Lab/FIM-Midtraining/blob/main/midtraining/configs/fim_midtrain.yaml) on [TIGER-Lab/FIM-Midtraining-400K](https://huggingface.co/datasets/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_8B.yaml`](https://github.com/TIGER-AI-Lab/FIM-Midtraining/blob/main/midtraining/configs/FIM_Midtrain_8B.yaml))
+- **Post-training**: none — see [TIGER-Lab/FIM-8B](https://huggingface.co/TIGER-Lab/FIM-8B) for the post-trained agent model
+
+## Serve with vLLM
+
+Ships the native Qwen3 40960 context (the yarn extension to 163840 was applied at post-training time); no overrides needed:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python -m vllm.entrypoints.openai.api_server \
+  --model TIGER-Lab/FIM-Mid-8B \
+  --served-model-name FIM-Mid-8B \
+  --host 127.0.0.1 \
+  --port 8400 \
+  --tensor-parallel-size 1 \
+  --max-model-len 40960 \
+  --gpu-memory-utilization 0.9 \
+  > vllm_fim_mid8b.log 2>&1 &
+```
+
+## Post-training
+
+To reproduce FIM-8B, run SWE-Lego trajectory SFT from this checkpoint — the exact config is [`posttraining/swe_lego/FIM_Posttrain_8B.yaml`](https://github.com/TIGER-AI-Lab/FIM-Midtraining/blob/main/posttraining/swe_lego/FIM_Posttrain_8B.yaml) (LLaMA-Factory, full fine-tuning, lr `1.0e-4`, **2 epochs** — the official SWE-Lego recipe's 4 overfits this base — cutoff 131072 with yarn rope scaling, `qwen3_nothink` template, `turn_mask` enabled), which already points at this repo id. See [`posttraining/swe_lego/`](https://github.com/TIGER-AI-Lab/FIM-Midtraining/tree/main/posttraining/swe_lego) 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}
+}
+```