Downloads · 30 days
5
11% of all-time downloads
QuanDS/legalir-bge-m3-retriever-v2-fold0
legalir-bge-m3-retriever-v2-fold0 is a machine learning model from QuanDS. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Vietnamese legal dense retriever fine-tuned from BAAI/bge-m3 for the UIT DSC 2026 LegalIR task. The training data contains 6,001 query-positive-negative groups from the shared retriever-v002 artifact. Document-group f…
Downloads · 30 days
5
11% of all-time downloads
All-time downloads
44
Public
Parameters
568M
2.3 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors2.3 GB · 99%
From the Hugging Face model README
Vietnamese legal dense retriever fine-tuned from BAAI/bge-m3 for the UIT DSC
2026 LegalIR task. The training data contains 6,001 query-positive-negative
groups from the shared retriever-v002 artifact. Document-group fold 0 (1,400
queries) was held out during training.
Training used one epoch, seven hard negatives per positive, CLS pooling,
normalized embeddings, BF16, learning rate 1e-5, and effective query batch
size 8. This repository contains the final model weights only; optimizer
checkpoints are intentionally omitted.
The held-out fold contains 1,535 gold document labels over 1,400 queries.
| Cutoff | Recall | Precision | F2 |
|---|---|---|---|
| 1 | 0.547143 | 0.573571 | 0.552232 |
| 2 | 0.730357 | 0.387143 | 0.620363 |
| 3 | 0.807917 | 0.286905 | 0.592664 |
| 5 | 0.881488 | 0.189857 | 0.509949 |
| 20 | 0.969345 | 0.053079 | 0.217712 |
| 100 | 0.987262 | 0.021907 | 0.100604 |
| 170 | 0.987262 | 0.021880 | 0.100490 |
These are retrieval-only diagnostics on the fullchunk-v011 corpus; they are not an end-to-end reranker score.
from FlagEmbedding import BGEM3FlagModel
model = BGEM3FlagModel("QuanDS/legalir-bge-m3-retriever-v2-fold0", use_fp16=True)
query_vecs = model.encode(["Câu hỏi pháp lý..."], return_dense=True)["dense_vecs"]