---
pipeline_tag: image-text-to-text
library_name: peft
base_model: google/gemma-3-4b-it
datasets:
  - Silviase/QuIC-360
license: gemma
tags:
  - cora
  - panorama
  - 360-degree
  - lora
  - vision-language
---

# CORA Gemma3-4B DenseCL 50%

Seed-42 LoRA adapter for the ECCV 2026 CORA release. It must be used with
`google/gemma-3-4b-it` at revision `093f9f388b31de276ce2de164bdc2081324b9767` and the matching CORA
config included as `cora_config.yaml`.

## Evaluation

QuIC-360 test set (5349 query-caption pairs):

| BLEU-4 | METEOR | ROUGE-L | CIDEr | SPICE |
|---:|---:|---:|---:|---:|
| 0.0445 | 0.1149 | 0.2455 | 0.3422 | 0.1667 |

Three-seed aggregate results and exact split hashes are maintained in the
[CORA-360 repository](https://github.com/wooseungw/CORA-360).

## Usage

```bash
git clone --branch v2.0.0-eccv2026 https://github.com/wooseungw/CORA-360.git
cd CORA-360
./reproduce.sh evaluate gemma3-4b-densecl /path/to/test.csv
```

## Limitations and license

CORA was evaluated on English query-focused captions from QuIC-360. Performance
outside that domain, on non-ERP imagery, or in safety-critical settings is not established.
This adapter is subject to the Gemma Terms of Use in addition to the CORA source-code license.

## Citation

```bibtex
@inproceedings{woo2026cora,
  title={Overlap-Consistent View Decomposition for Adapting Vision--Language Models to 360-Degree Panoramas},
  author={Woo, Seungwoo and Jung, Daewon and Youm, Sekyoung},
  booktitle={European Conference on Computer Vision},
  year={2026}
}
```