FIM-14B

📄 Paper · 💻 GitHub · 🤗 Dataset · 🤗 Collection

FIM-14B is the 14B coding-agent model of "Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models": Qwen2.5-Coder-14B-Instruct, mid-trained on function-aware FIM data, then post-trained on R2E-Gym agent trajectories with the upstream recipe unmodified. The mid-training stage is the only difference from a standard R2E-Gym reproduction — worth +3.0 points on SWE-Bench-Verified and +4.0 on SWE-Bench-Lite, while also recovering most of the general-capability erosion that agentic post-training inflicts (LiveCodeBench +11.1, τ-bench +3.9, BFCL +2.4 over the post-training-only arm).

Training pipeline

• Base model: Qwen/Qwen2.5-Coder-14B-Instruct
• FIM mid-training: midtraining/configs/fim_midtrain.yaml on TIGER-Lab/FIM-Midtraining-400K (as-run copy: FIM_Midtrain_14B.yaml) → intermediate checkpoint released as TIGER-Lab/FIM-Mid-14B
• Post-training: SFT on R2E-Gym agent trajectories — posttraining/r2egym/ (as-run copy: FIM_Posttrain_14B.yaml)

Results

Means over three evaluation seeds, identical harness for both arms (paper Tables 1–2):

Setting	SWE-Bench-Verified	SWE-Bench-Lite
Qwen2.5-Coder-14B-Instruct + R2E-Gym (reproduced)	26.20	18.00
FIM-14B (+ FIM mid-training)	29.20	22.00
Δ	+3.00	+4.00

Capability preservation at 14B (six benchmarks outside SWE-Bench):

Setting	LiveCode	OJBench	FSB-EN	Terminal	τ-bench	BFCL	Avg
+ R2E-Gym only	24.10	2.80	47.72	2.41	3.40	15.80	16.04
FIM-14B	35.20	4.74	48.25	3.66	7.30	18.20	19.56

Reproduction guides for all six: evaluation/.

Evaluate on SWE-Bench Verified

FIM-14B is evaluated with the R2E-Gym agent scaffold (fixed by its post-training pipeline). The complete pinned walkthrough lives at evaluation/swebench/released_checkpoints.md.

1. Serve the model with vLLM

CUDA_VISIBLE_DEVICES=0 \
VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
python -m vllm.entrypoints.openai.api_server \
  --model TIGER-Lab/FIM-14B \
  --served-model-name FIM-14B \
  --host 127.0.0.1 \
  --port 8400 \
  --tensor-parallel-size 1 \
  --max-model-len 65536 \
  --hf-overrides '{"max_position_embeddings": 65536}' \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.9 \
  > vllm_fim14b.log 2>&1 &

Wait until the server is up (model load takes ~1 minute):

curl -s http://127.0.0.1:8400/v1/models

2. Run the agent on SWE-Bench Verified

From an upstream, unmodified R2E-Gym checkout (Docker required):

export OPENAI_API_KEY=EMPTY
export LLM_BASE_URL="http://127.0.0.1:8400/v1"

uv run python src/r2egym/agenthub/run/edit.py runagent_multiple \
  --dataset "R2E-Gym/SWE-Bench-Verified" \
  --split "test" \
  --start_idx 0 \
  --k 500 \
  --traj_dir "./traj" \
  --exp_name "FIM-14B_swebench_verified_r1" \
  --llm_name "openai/FIM-14B" \
  --scaffold "r2egym" \
  --backend "docker" \
  --use_fn_calling False \
  --temperature 0 \
  --max_steps 40 \
  --max_steps_absolute 100 \
  --max_workers 6 \
  --max_reward_calc_time 1200 \
  --max_tokens 65536 \
  --use_existing True

For SWE-Bench Lite, use --dataset "R2E-Gym/SWE-Bench-Lite" --k 300.

3. Score with the official SWE-bench harness

Convert the trajectories to a submission and score with the official harness — evaluation/swebench/score.sh. The reported number is resolved_instances / total_instances.

Citation

@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}
}