internvl3_5-4b-iconclass-sft-brillfull

LoRA adapter for OpenGVLab/InternVL3_5-4B-HF fine-tuned with ms-swift on davanstrien/iconclass-vlm-brillfull (86216 training rows, config sft).

epochs	1.0
learning rate	0.0001
LoRA rank / alpha	8 / 32
effective batch	16
max length / pixels	2560 / 1003520
vision tower	frozen
final train loss	0.48287477493286135
eval loss	0.4837830364704132
training time	1023.1 min

Use

swift infer --model OpenGVLab/InternVL3_5-4B-HF --adapters davanstrien/internvl3_5-4b-iconclass-sft-brillfull --use_hf true --stream true

Reproduction

Produced on Hugging Face Jobs (gpu) with the swift-vlm-sft.py recipe from uv-scripts. Run it yourself:

hf jobs uv run --flavor a10g-large --timeout 1h --secrets HF_TOKEN https://huggingface.co/datasets/uv-scripts/finetune/raw/main/swift-vlm-sft.py davanstrien/iconclass-vlm-brillfull davanstrien/internvl3_5-4b-iconclass-sft-brillfull --model OpenGVLab/InternVL3_5-4B-HF --config sft --epochs 1 --max-tiles 6 --max-length 2560 --batch-size 2 --grad-accum 8 --eval-ratio 0.005 --save-steps 500 --work-dir /data/iconclass-internvl-full-v2