---
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.