AcciSAGE-S8c-SAPO-v2-ckpt75
This repository contains a LoRA adapter checkpoint for AcciSAGE spatial grounding experiments on traffic accident videos.
Model
- Base model:
Qwen3-VL-8B-Instruct
- Local base model path used in training:
models/Qwen3-VL-8B-Instruct
- Adapter type: LoRA
- Task: accident spatial grounding
- Output format:
{"accident_center": [x, y]}
- Coordinate system: relative_1000_xy
- Input setting: oracle-time 8-second accident clip
Training stage
- Stage: S8c Stage C SAPO calibration
- Variant:
v2
- Reward: continuous piecewise-linear reward v2; stricter Hit@150-oriented calibration
- Dataset: S8c-SAPO curated280
- Initialization: S8c Stage B checkpoint-260
Eval200 summary
| Metric | Value |
|---|
| parse_ok_rate | 1.0 |
| valid_coord_rate | 1.0 |
| mean L2 | 139.7273 |
| median L2 | 94.9230 |
| P75 L2 | 177.2763 |
| P90 L2 | 296.6735 |
| badcase rate, d > 150 | 0.3350 |
This repository contains only the adapter checkpoint, not the base Qwen3-VL model weights.
