---
license: mit
library_name: pytorch
tags: [visual-question-answering, vivqa, beit3, vietnamese, multimodal]
language: [vi]
---

# ViVQA-BEiT3 — Expert A (clean baseline)

BEiT-3 **base** (patch16, 224) finetuned on **ViVQA** (Vietnamese VQA, 218 answer classes).
This is the **clean baseline** expert (no augmentation) from the question-type routing study.

## Results (test, 3001 samples)

| Total | what/other | counting | color | location |
|------|-----------|----------|-------|----------|
| **70.21%** | 71.53 | 61.26 | 78.40 | 66.13 |

## Training
- Init: `beit3_base_indomain_patch16_224`, XLM-R (`beit3.spm`) tokenization
- 30 epochs, lr 5e-5, layer_decay 0.8, batch 64, seed 42, `--clip_grad 1.0`, `--randaug`
- `--augmented none`

## Files
- `checkpoint-best.pth` — best-val checkpoint (`model` key in state_dict)

## Load
```python
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("ThucPD/vivqa-beit3-routing-A-clean", "checkpoint-best.pth")
# use with BEiT-3 run_beit3_finetuning.py --model beit3_base_patch16_224 --finetune <ckpt>
```

See the routing study: per-type augmentation effects reproduce but are sub-noise; this clean
baseline is the robust optimum. Paired with the YOLO-augmented expert `ThucPD/vivqa-beit3-routing-C-yolo`.