algocean-executor / README.md
Algocean's picture
Initial public release (renamed from clx- prefix to algocean-)
d2cedd3 verified
|
Raw
History Blame Contribute Delete
1.92 kB
metadata
license: apache-2.0
language:
  - ko
  - en
task_categories:
  - text-generation
  - question-answering
tags:
  - lora
  - sft
  - rag
  - grounded-generation
  - abstention
  - korean
size_categories:
  - 100K<n<1M
pretty_name: algocean-executor

algocean-executor

질의와 근거를 받아 답을 쓰거나, 근거가 부족하면 기권하도록 가르치는 LoRA SFT 데이터셋입니다.
파이프라인에서 최종 답변을 쓰는 노드입니다.

규모

파일 행 수
executor.jsonl 150,000
executor.eval.jsonl 2,000
  • 형식: JSONL, messages 3턴
  • 언어: 한국어 약 70% · 영어 약 30%

어디에 쓰나요

  • 주어진 근거만으로 답하는 grounded generation
  • 근거 없음 / 충돌 / 개인 메모뿐일 때 기권·조건부 답변
  • 사내 QA에서 “근거 없이 단정하지 않기”

어떤 모델에 LoRA 하나요

  • 베이스: 1B ~ 9B 인스트럭트 모델
  • rank 816, alpha 1632, lr 1e-42e-4, epochs 23
  • loss: assistant 턴만
  • 장문 tier는 컨텍스트에 맞게 필터

사용

from datasets import load_dataset

ds = load_dataset("Algocean/algocean-executor", data_files="executor.jsonl", split="train")
ev = load_dataset("Algocean/algocean-executor", data_files="executor.eval.jsonl", split="train")
  • 학습: messages
  • 메타는 필터용

assistant 출력 (요약)

{
  "answer": "...",
  "evidence_refs": [{"document_id": "...", "span": "..."}],
  "abstain": false,
  "conditional": false,
  "reason_code": "SUPPORTED"
}

reason_code: SUPPORTED / NO_EVIDENCE / PERSONAL_ONLY / CONFLICTING
conditional=true이면 답은 쓰되 성립 조건을 밝힙니다.

식별자·문서 id는 중립 코드(mNN, doc-NNNN, Brand-* 등)로 정규화되어 있다. 학습에는 messages만 쓴다.

라이선스

Apache-2.0