CoolFace
Datasetpublic

lius-cc/daoism-sft-v1

Daoism SFT Dataset v1 道教 LLM SFT 訓練資料集 — Phase 1-2 蒸餾+清洗產物。配 Qwen3.6-27B QLoRA 微調用。 資料來源 從 lius-cc 道教知識庫(vault: concept/deity/scripture/ritual/paper/person)蒸餾。生成模型: ChatGPT gpt-5.5-mini(主力) Gemini 2.5 Fast / Flash(多源整合) Pipeline Phase 1 蒸餾:vault → instruction/response pair(578k raw) Phase 2 清洗: critic_score ≥ 80 過濾 長度過濾(instr ≥10 字、resp 80-2000 字) 精確去重 + 語意去重(gemini-embedding-2, cos < 0.97) 紅線檢查(醫療/命名/墓地/法事金額)→ 加 disclaimer 邊緣 pair 改寫(80-85 → 90+… See the full description on the dataset page: https://huggingface.co/datasets/lius-cc/daoism-sft-v1.

sourceHugging Faceupdated 5mo agoView on Hugging Face
0likes16downloads
5 commits on main
83e2a945mo ago

phase 2 cleanup 2026-05-14

chiyingliu
cae2c4e5mo ago

rollback bad phase 2 dedup output

chiyingliu
3e0c23f5mo ago

rollback bad phase 2 dedup output

chiyingliu
14813cd5mo ago

phase 2 cleanup 2026-05-14

chiyingliu
3747a965mo ago

initial commit

chiyingliu