surgvidlm-heico-stage2

SurgVidLM HEICO Stage 2

This repository contains the merged, standalone HEICO stage-2 model for SurgVidLM.

Model

• Base model: SurgVidLM/surgvidlm/surgvidlm-stage1
• Fine-tuning method: LoRA
• Selected adapter checkpoint: LLaMA-Factory/saves/lora/heico_frame_stage2/checkpoint-800
• Exported merged model: LLaMA-Factory/saves/model/heico-frame-stage2

Task

Frame-level question answering on the HEICO dataset.

Training setup

• Training script: train_heico_frame_stage2.sh
• Training config: LLaMA-Factory/examples/train_lora/heico_frame_stage2_lora_sft.yaml
• Adapter training output: LLaMA-Factory/saves/lora/heico_frame_stage2

Evaluation

Validation run used the selected LoRA adapter on the HEICO validation split. The merged model was also evaluated on the official HEICO test split.

Usage

Run inference with the merged model:

./infer_heico_frame_stage2.sh merged \
  /home/ah445458/crcv/SurgVidLM/LLaMA-Factory/saves/model/heico-frame-stage2 \
  results/heico_test.json

Run ORENA evaluation on predictions:

./infer_heico_frame_stage2.sh evaluate-orena \
  results/heico_test.json \
  results/heico_test_orena

Notes

• This is a frame-native stage-2 model.
• It does not use the original full-video fusion path.
• The merged model is the easiest artifact to upload and use for inference.