hoailebads's picture
Add VLSP legal reranker train (9994) + eval (296) data + dataset card
65a4d8e verified
|
Raw
History Blame Contribute Delete
3.83 kB
---
license: apache-2.0
language:
- vi
task_categories:
- text-ranking
tags:
- reranker
- cross-encoder
- legal
- vietnamese
- vlsp
- hard-negatives
pretty_name: VLSP Vietnamese Legal Reranker Train + Eval
size_categories:
- 1K<n<10K
configs:
- config_name: train
data_files: train.jsonl
default: true
- config_name: eval
data_files: eval.json
---
# VLSP Vietnamese Legal Reranker — Training + Eval data
Bộ dữ liệu **listwise reranking** cho truy hồi điều luật tiếng Việt (VLSP). Đây **chính xác**
dữ liệu đã dùng để fine-tune 2 model reranker:
- 🤗 [`hoailebads/Qwen3-Reranker-8B-VLSP-Legal-LoRA`](https://huggingface.co/hoailebads/Qwen3-Reranker-8B-VLSP-Legal-LoRA) (bản chính, vượt SOTA VLSP)
- 🤗 [`hoailebads/Qwen3-Reranker-0.6B-VLSP-Legal-LoRA`](https://huggingface.co/hoailebads/Qwen3-Reranker-0.6B-VLSP-Legal-LoRA) (bản nhẹ)
📦 Code, pipeline & script tạo dữ liệu: **https://github.com/hoaileba/Qwen-Retrieval-Tuning**
(file gốc: `dataset/train_rankllama_hn23_v3.jsonl`, tạo bằng `data_prep/mine_hard_negatives_v3.py`).
---
## Nội dung
| File | Split | Số dòng | Định dạng | Mô tả |
|---|---|---|---|---|
| `train.jsonl` | train | **9,994** | JSONL | mỗi dòng = 1 query + 1 positive + tối đa 23 hard negative |
| `eval.json` | eval | **296** | JSON array | query + 1 positive (dedup theo `aid` còn **219** query unique) |
### `train.jsonl` — schema mỗi dòng
```json
{
"id": "d9189b1cc63747b4e879984e5a72d190",
"question": "Những trường hợp nào không tiếp tục sử dụng đối với công trình hết thời hạn...",
"positive": {"id": "06/2021/nđ-cp|41", "text": "Điều 41. Xử lý đối với công trình..."},
"negatives": ["Điều 19. Vi phạm quy định về bảo hành...", "..."],
"_source": "zalo",
"_neg_source": "teacher_aiteamvn_v3"
}
```
- `positive`**điều luật đúng** (index 0 khi train listwise).
- `negatives`**hard negative** mine bằng dense retriever + teacher reranker (23/query cho
phần lớn mẫu).
- Loss listwise: positive ở vị trí 0, negatives là các lựa chọn còn lại (InfoNCE, không dùng
in-batch negative).
### `eval.json` — schema mỗi record
```json
{"id": "11075", "question": "Người đi xe máy đâm vào xe ôtô...", "positive": {"id": 56886, "text": "Luật quy định về giao thông..."}}
```
Query eval **chỉ có ground-truth positive**; ứng viên để rerank được sinh riêng bằng bi-encoder
(top-100), xem repo GitHub.
---
## Thống kê `train.jsonl`
| Thuộc tính | Giá trị |
|---|---|
| Tổng số mẫu | 9,994 |
| Nguồn (`_source`) | `synthetic` 5,530 · `zalo` 2,477 · `vlsp` 1,987 |
| Nguồn negative (`_neg_source`) | `teacher_aiteamvn_v3` (toàn bộ) |
| Số negative/mẫu | 23 (9,051 mẫu) · 19 (59) · các mức thấp hơn cho phần còn lại |
## Cách nạp
```python
from datasets import load_dataset
train = load_dataset("hoailebads/VLSP-Legal-Reranker-Train", "train", split="train")
eval_ = load_dataset("hoailebads/VLSP-Legal-Reranker-Train", "eval", split="train")
print(train[0]["question"], train[0]["positive"]["id"], len(train[0]["negatives"]))
```
## Giới hạn
- Miền hẹp: **điều luật tiếng Việt** (VLSP + Zalo legal). Không dùng cho miền khác mà không re-mine
hard negative.
- 55% mẫu train là `synthetic` (câu hỏi sinh tự động) — xem thảo luận về overfit/nhãn trong
[`results/`](https://github.com/hoaileba/Qwen-Retrieval-Tuning/tree/main/results) của repo.
- Hard negative là **teacher-mined**, có thể chứa false negative (điều luật thực ra liên quan).
## License
Apache-2.0. Văn bản luật gốc thuộc phạm vi công (VLSP / Zalo legal corpus).