| base_model: google/gemma-3-4b-it | |
| library_name: peft | |
| tags: | |
| - gui-grounding | |
| - screenspot | |
| - portal | |
| - lora | |
| license: mit | |
| # portal-vlm: gemma3-lora-1k | |
| Fresh-training reference for the cross-family porting row: grounding LoRA (r16/a32, LM sites) on Gemma-3-4B-it at ~1k examples, canonical 0-1000 dialect. | |
| Part of the [portal-vlm](https://github.com/robbym-dev/portal-vlm) release - an | |
| independent replication of Ramp Labs' PorTAL (portable task adapters via | |
| hypernet-generated LoRA) extended to vision-language models on GUI grounding. | |
| | | ScreenSpot-v2 overall | web split | | |
| |---|---|---| | |
| | this artifact | **15.8%** | 10.1% | | |
| - Base model: `google/gemma-3-4b-it` @ `093f9f388b31de276ce2de164bdc2081324b9767` | |
| - Training config: [`configs/gemma3_lora_1k.yaml`](https://github.com/robbym-dev/portal-vlm/blob/main/configs/gemma3_lora_1k.yaml) | |
| - Eval record: [`results/gemma3_lora_1k.json`](https://github.com/robbym-dev/portal-vlm/blob/main/results/gemma3_lora_1k.json) | |
| ## Reproduce this row without training | |
| ```bash | |
| git clone https://github.com/robbym-dev/portal-vlm && cd portal-vlm && uv sync | |
| uv run python scripts/eval.py --config configs/gemma3_lora_1k.yaml --adapter hf:manihani4/portal-vlm-gemma3-lora-1k | |
| ``` | |
| Standard PEFT LoRA adapter - also loadable directly with `peft.PeftModel.from_pretrained` on the pinned base model. | |