FYP2526-HC2/smartsearch-qwen3-vl-embedding-2b-lora-ckpt9905

SmartSearch Qwen3-VL-Embedding-2B LoRA (checkpoint-9905)

This repository contains an inference-focused LoRA adapter exported from local training checkpoint checkpoint-9905.

Model Details

• Base model: Qwen/Qwen3-VL-Embedding-2B
• Adapter type: LoRA (PEFT)
• Primary use: medical image embedding and image-label matching (EI vs ET)
• Training framework: Swift + PEFT (see args.json)

Files

• adapter_model.safetensors: LoRA adapter weights
• adapter_config.json: LoRA configuration
• additional_config.json: extra runtime config
• args.json: training/eval arguments snapshot

Evaluation Snapshot (checkpoint-9905)

Metrics below are from instruction-sweep best results in local logs (results/eiet_cls_instruction_sweep_ckpt9905_best_only):

Dataset	Accuracy	F1	AUC
D010_RSNA	0.7162	0.6553	0.7922
D027_NLMTB	0.5375	0.1190	0.8223
D037_WCE	0.7712	0.7699	0.9598
D043_UBIBC	0.5050	0.0050	0.6227
D046_BUSBRA	0.6812	0.0000	0.5651
D136_ChestXRay2017	0.3750	0.0000	0.6573
D153_OCTDL	0.0813	0.0275	0.6422

Quick Start

from transformers import AutoModel, AutoProcessor
from peft import PeftModel

base_id = "Qwen/Qwen3-VL-Embedding-2B"
adapter_id = "Fred25022004/smartsearch-qwen3-vl-embedding-2b-lora-ckpt9905"

processor = AutoProcessor.from_pretrained(base_id, trust_remote_code=True)
base_model = AutoModel.from_pretrained(base_id, trust_remote_code=True)
model = PeftModel.from_pretrained(base_model, adapter_id)

Intended Use

• Research and benchmarking for medical image embedding/classification retrieval-style tasks.

Limitations

• This adapter is domain-tuned and may not generalize well outside medical imagery.
• Performance varies significantly by dataset/class balance.

License

This adapter should follow the base model terms and the dataset usage terms used in training/evaluation. Please verify compatibility before commercial use.