---
language: vi
tags:
  - dpr
  - dense-passage-retrieval
  - phobert
  - vietnamese
  - question-answering
  - retrieval
license: mit
---

# HUMG DPR Model

DPR Dual Encoder fine-tuned from `vinai/phobert-base` for HUMG regulation retrieval.

## Current uploaded checkpoint

- Local source: `inferences/dpr_ckpt_projection_trial/model.pt`
- Training mode: full-softmax projection trial over all passages
- Export note: `model.pt` is uploaded both as `model.pt` and `model_last.pt` so existing runtimes that load `model_last.pt` use the best checkpoint.

## Test metrics on `data/splits/test.json`

- Recall@1: 0.5031
- Recall@3: 0.7739
- Recall@5: 0.8609
- Recall@10: 0.9242
- Recall@20: 0.9429
- Recall@50: 0.9689
- MRR: 0.6513

The FAISS artifacts in `mudotet/humg-search-data` were rebuilt from this checkpoint.