This adapter estimates the first clear impact time in traffic accident videos. It was trained for the ACCIDENT-style traffic accident video understanding setting.

Inference example

bash

export IMAGE_MAX_TOKEN_NUM=1024
export VIDEO_MAX_TOKEN_NUM=512
export FPS_MAX_FRAMES=128
export FPS=4
export FORCE_QWENVL_VIDEO_READER=torchcodec

swift infer \
  --model Qwen/Qwen3-VL-8B-Instruct \
  --adapters <this_repo_or_local_adapter_path> \
  --val_dataset path/to/infer.jsonl \
  --max_new_tokens 64 \
  --temperature 0 \
  --result_path infer_results.jsonl

yzk9/accident-vlm-t8-sapo-v4b-qwen3vl8b-lora

Accident-VLM T8-SAPO-v4b Qwen3-VL-8B LoRA Adapter

Base model

Intended use

Inference example

yzk9/
accident-vlm-t8-sapo-v4b-qwen3vl8b-lora