Initial public release (renamed from clx- prefix to algocean-)
Browse files- .gitattributes +1 -0
- README.md +81 -0
- executor.eval.jsonl +0 -0
- executor.jsonl +3 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
executor.jsonl filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- ko
|
| 5 |
+
- en
|
| 6 |
+
task_categories:
|
| 7 |
+
- text-generation
|
| 8 |
+
- question-answering
|
| 9 |
+
tags:
|
| 10 |
+
- lora
|
| 11 |
+
- sft
|
| 12 |
+
- rag
|
| 13 |
+
- grounded-generation
|
| 14 |
+
- abstention
|
| 15 |
+
- korean
|
| 16 |
+
size_categories:
|
| 17 |
+
- 100K<n<1M
|
| 18 |
+
pretty_name: algocean-executor
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# algocean-executor
|
| 22 |
+
|
| 23 |
+
질의와 근거를 받아 **답을 쓰거나, 근거가 부족하면 기권**하도록 가르치는 LoRA SFT 데이터셋입니다.
|
| 24 |
+
파이프라인에서 **최종 답변을 쓰는 노드**입니다.
|
| 25 |
+
|
| 26 |
+
## 규모
|
| 27 |
+
|
| 28 |
+
| 파일 | 행 수 |
|
| 29 |
+
|------|------:|
|
| 30 |
+
| `executor.jsonl` | 150,000 |
|
| 31 |
+
| `executor.eval.jsonl` | 2,000 |
|
| 32 |
+
|
| 33 |
+
- 형식: JSONL, `messages` 3턴
|
| 34 |
+
- 언어: 한국어 약 70% · 영어 약 30%
|
| 35 |
+
|
| 36 |
+
## 어디에 쓰나요
|
| 37 |
+
|
| 38 |
+
- 주어진 근거만으로 답하는 grounded generation
|
| 39 |
+
- 근거 없음 / 충돌 / 개인 메모뿐일 때 **기권·조건부 답변**
|
| 40 |
+
- 사내 QA에서 “근거 없이 단정하지 않기”
|
| 41 |
+
|
| 42 |
+
## 어떤 모델에 LoRA 하나요
|
| 43 |
+
|
| 44 |
+
- 베이스: **1B ~ 9B** 인스트럭트 모델
|
| 45 |
+
- rank 8~16, alpha 16~32, lr 1e-4~2e-4, epochs 2~3
|
| 46 |
+
- loss: assistant 턴만
|
| 47 |
+
- 장문 tier는 컨텍스트에 맞게 필터
|
| 48 |
+
|
| 49 |
+
## 사용
|
| 50 |
+
|
| 51 |
+
```python
|
| 52 |
+
from datasets import load_dataset
|
| 53 |
+
|
| 54 |
+
ds = load_dataset("Algocean/algocean-executor", data_files="executor.jsonl", split="train")
|
| 55 |
+
ev = load_dataset("Algocean/algocean-executor", data_files="executor.eval.jsonl", split="train")
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
- 학습: `messages` 만
|
| 59 |
+
- 메타는 필터용
|
| 60 |
+
|
| 61 |
+
### assistant 출력 (요약)
|
| 62 |
+
|
| 63 |
+
```json
|
| 64 |
+
{
|
| 65 |
+
"answer": "...",
|
| 66 |
+
"evidence_refs": [{"document_id": "...", "span": "..."}],
|
| 67 |
+
"abstain": false,
|
| 68 |
+
"conditional": false,
|
| 69 |
+
"reason_code": "SUPPORTED"
|
| 70 |
+
}
|
| 71 |
+
```
|
| 72 |
+
|
| 73 |
+
`reason_code`: `SUPPORTED` / `NO_EVIDENCE` / `PERSONAL_ONLY` / `CONFLICTING`
|
| 74 |
+
`conditional=true`이면 답은 쓰되 성립 조건을 밝힙니다.
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
식별자·문서 id는 중립 코드(`mNN`, `doc-NNNN`, `Brand-*` 등)로 정규화되어 있다. 학습에는 `messages`만 쓴다.
|
| 78 |
+
|
| 79 |
+
## 라이선스
|
| 80 |
+
|
| 81 |
+
Apache-2.0
|
executor.eval.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
executor.jsonl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3b27502fea0f8bcef5a6b76626fdcf2266c9499f2dec8c302a3d550243370836
|
| 3 |
+
size 239099813
|