0927_v2

Qwen3.5-9B full-parameter SFT checkpoint-140, exported to BF16 for inference. This is a complete model, not a LoRA adapter. The original FP32 training checkpoint remains on the training server.

Training: 430 new-data training examples and all 1455 July historical annotation rows; 8 A800 GPUs, global batch 8, gradient accumulation 1, warmup 0, AdamW, learning rate 1e-5, cosine schedule, 200 optimizer steps. The fixed 92-image new-data validation set was used to choose checkpoint-140. The 93-image fixed test split was not used for this selection.

Recall-first validation (92 positive images): macro image-class recall 0.68164, micro image-class recall 0.78462. Strict type-correct localized instance matches (IoU ≥ 0.5): 42/130. The validation split has no normal images, so normal false-alarm rate cannot be measured. These numbers do not establish deployment readiness.

Use inference_prompt.txt as the user text accompanying each image. Output coordinates are 0–1000 normalized integers. Inference tested using BF16, greedy decoding, max_new_tokens=192. The result is for visual defect detection; seeing a person alone does not prove unauthorized intrusion.

The corresponding dataset is in the private repository ZJUxxc/0927_training_data (access required).