Datasets:
Kinshield SMS — 20260917
Artifacts SFT mới nhất: train của campaign Sonnet corrective284-r01 và hai tập OOD theo test-only gold revision label_review001.
| Split | Requests | Parents |
|---|---|---|
| train | 77,358 | 25,786 |
| test_ood_scenario | 384 | 128 |
| test_ood_entity | 384 | 128 |
Hai tập OOD đều có 192 mẫu VI và 192 mẫu EN. Không trùng request ID, parent ID hoặc input sms+evidence giữa ba splits. Dev không nằm trong gói này theo phạm vi export.
Contract
Giữ nguyên JSONL và ID/index gốc. input chỉ gồm sms và evidence, mỗi evidence có subject. output gồm verdict, key_findings, recommendations. Các trường khác là metadata; không đưa gold hoặc ID nguồn vào classifier prompt.
Dùng system_prompt.txt làm system message, JSON serialize row['input'] làm user message và JSON serialize row['output'] làm assistant target. Catalog IDs phải giữ chính xác; output_schema.json mô tả output. Không có NER như input riêng; bundle này không thay thế raw artifacts NER upstream.
Đọc JSONL bằng thư viện chuẩn để bảo toàn evidence có cấu trúc khác nhau theo tool:
import json
from huggingface_hub import hf_hub_download
path = hf_hub_download('jacky-qualgo/kinshield-sms-20260917',
'data/train.jsonl', repo_type='dataset')
with open(path, encoding='utf-8') as f:
for line in f:
row = json.loads(line)
- Downloads last month
- 18