kissin42 commited on
Commit
a873bcd
·
verified ·
1 Parent(s): af9985d

fix: replace attached '~' (approx) with ≈ so HF markdown stops rendering it as strikethrough; backtick 'T ~ U' notation

Browse files
Files changed (1) hide show
  1. README.ko.md +86 -86
README.ko.md CHANGED
@@ -1,86 +1,86 @@
1
- ---
2
- license: cc-by-4.0
3
- tags:
4
- - reinforcement-learning
5
- - unity-ml-agents
6
- - minari
7
- - causal-gpt-rl
8
- ---
9
-
10
- # Causal GPT-RL — Unity ML-Agents 궤적
11
-
12
- 모델이 제거된 Unity 빌드를 스톡 정책으로 외부 제어해 수집한 궤적을
13
- [Minari](https://minari.farama.org) 0.5.3 형식으로 제공합니다. 공개된 환경,
14
- 데이터셋, 스톡 정책은 모두 **ML-Agents `release_23`** 기준입니다.
15
-
16
- ## 데이터셋
17
-
18
- | Dataset id | Episodes | Transitions | Observation | Action |
19
- |---|---:|---:|---|---|
20
- | `unity/crawler/expert-v0` | 1,048 | 1,000,002 | `Tuple(Box(126), Box(32))` | `Box(20)` |
21
- | `unity/pushblock/expert-v0` | 52,801 | 1,000,000 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
22
- | `unity/soccer-twos/expert-v0` | 14,588 | 1,000,816 | ego `Dict(Tuple(Box(264), Box(72)))` | ego `Dict(MultiDiscrete([3,3,3]))` |
23
- | `unity/dungeon-escape/expert-v0` | 35,787 | 1,003,523 | ego `Dict(Tuple(Box(10), Box(360), Box(1)))` | ego `Dict(Discrete(7))` |
24
- | `unity/3dball-hard/expert-v0` | 1,009 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
25
- | `unity/pyramids/expert-v0` | 5,348 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
26
- | `unity/worm/expert-v0` | 1,000 | 1,000,000 | `Box(64)` | `Box(9)` |
27
- | `unity/walker/expert-v0` | 1,458 | 1,000,010 | `Box(243)` | `Box(39)` |
28
-
29
- 각 데이터셋은 `<환경>/<tier>/data/main_data.hdf5`와 `metadata.json`으로
30
- 구성됩니다. 관측 센서는 의미가 섞이지 않도록 별도의 `Tuple` leaf로 보존하며,
31
- 센서가 하나인 환경은 단일 `Box`를 사용합니다.
32
-
33
- SoccerTwos와 DungeonEscape는 공유 정책을 사용하는 분산형 다중 에이전트
34
- 환경입니다. 각 물리 에이전트는 ego episode 하나로 저장되며, 같은 경기나 협동
35
- 그룹은 함께 제공되는 match/group metadata로 연결됩니다.
36
-
37
- DungeonEscape·SoccerTwos·Pyramids·Crawler·PushBlock·3DBallHard·Worm·Walker는
38
- `expert` 외에 스톡 정책을 점진적으로 열화시킨 `medium`/`simple` 품질 래더
39
- 티어도 제공합니다. 티어는 Minari 정규화 스킬 기준(`simple` ~0.60, `medium` ~0.80,
40
- `expert` = 1.0)으로 보정됩니다. 열화 방식은 씬에 따라 다릅니다(seed 2310000) —
41
- 연속 제어 씬은 에피소드마다 보정된 가우시안 노이즈 범위를 씁니다. 단일 에이전트
42
- 목표 게임인 PushBlock과 Pyramids는 정책 자체 로짓에 대한 softmax temperature
43
- `T`(`softmax(logits/T)`)를 보정된 범위에서 **에피소드마다** 샘플링(각 에피소드가
44
- 자체 스킬 레벨)합니다. 협동 DungeonEscape은 경기마다 보정된
45
- 범위에서 softmax temperature `T`를 **그룹 단위로** 샘플링해(그룹의 세 에이전트가
46
- 같은 `T`를 공유 — 한 시점의 초기 체크포인트처럼 하나의 일관된 스킬 레벨) 역시
47
- step-reward로 채점하며, 자기대전 SoccerTwos는 매치마다 팀별로 softmax temperature
48
- `T`를 (더 넓은) 보정 범위에서 **독립 샘플링**합니다(두 팀이 독립적으로 뽑아 매치마다
49
- 실력차가 생김 — 여러 체크포인트가 붙는 리그전처럼). 목표
50
- 게임은 에피소드 리턴이 이봉(해결 vs 타임아웃)이라 **step-reward**(리턴/길이)로
51
- 채점합니다. 각 티어 평균은 목표에 맞으면서 에피소드들이 스킬 스펙트럼을 연속적으로
52
- 덮습니다. 티어별 수치는 영문 카드(README.md)의 per-environment quality ladder
53
- 표를 참고하세요.
54
-
55
- ## 동반 저장소
56
-
57
- - 모델 제거 Unity 빌드와 스톡 ONNX:
58
- [ccnets/causal-gpt-rl-unity-envs](https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-envs)
59
- - 학습된 Causal GPT-RL 정책:
60
- [ccnets/causal-gpt-rl-unity](https://huggingface.co/ccnets/causal-gpt-rl-unity)
61
- - 수집 및 Minari 패키징 코드:
62
- [examples/unity_collection](https://github.com/ccnets-team/causal-gpt-rl/tree/main/examples/unity_collection)
63
-
64
- ## 로드 예시
65
-
66
- ```python
67
- from pathlib import Path
68
- from huggingface_hub import snapshot_download
69
- import minari
70
-
71
- root = Path.home() / ".minari" / "datasets" / "unity"
72
- snapshot_download(
73
- repo_id="ccnets/causal-gpt-rl-unity-datasets",
74
- repo_type="dataset",
75
- allow_patterns="crawler/**",
76
- local_dir=root,
77
- )
78
- dataset = minari.load_dataset("unity/crawler/expert-v0")
79
- print(dataset.total_episodes, dataset.total_steps)
80
- ```
81
-
82
- ## 라이선스
83
-
84
- 궤적 데이터는 Creative Commons Attribution 4.0 International
85
- (CC-BY-4.0)로 배포됩니다. *ccnets — Causal GPT-RL*과 Unity ML-Agents 원본
86
- 환경을 표기해 주세요.
 
1
+ ---
2
+ license: cc-by-4.0
3
+ tags:
4
+ - reinforcement-learning
5
+ - unity-ml-agents
6
+ - minari
7
+ - causal-gpt-rl
8
+ ---
9
+
10
+ # Causal GPT-RL — Unity ML-Agents 궤적
11
+
12
+ 모델이 제거된 Unity 빌드를 스톡 정책으로 외부 제어해 수집한 궤적을
13
+ [Minari](https://minari.farama.org) 0.5.3 형식으로 제공합니다. 공개된 환경,
14
+ 데이터셋, 스톡 정책은 모두 **ML-Agents `release_23`** 기준입니다.
15
+
16
+ ## 데이터셋
17
+
18
+ | Dataset id | Episodes | Transitions | Observation | Action |
19
+ |---|---:|---:|---|---|
20
+ | `unity/crawler/expert-v0` | 1,048 | 1,000,002 | `Tuple(Box(126), Box(32))` | `Box(20)` |
21
+ | `unity/pushblock/expert-v0` | 52,801 | 1,000,000 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
22
+ | `unity/soccer-twos/expert-v0` | 14,588 | 1,000,816 | ego `Dict(Tuple(Box(264), Box(72)))` | ego `Dict(MultiDiscrete([3,3,3]))` |
23
+ | `unity/dungeon-escape/expert-v0` | 35,787 | 1,003,523 | ego `Dict(Tuple(Box(10), Box(360), Box(1)))` | ego `Dict(Discrete(7))` |
24
+ | `unity/3dball-hard/expert-v0` | 1,009 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
25
+ | `unity/pyramids/expert-v0` | 5,348 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
26
+ | `unity/worm/expert-v0` | 1,000 | 1,000,000 | `Box(64)` | `Box(9)` |
27
+ | `unity/walker/expert-v0` | 1,458 | 1,000,010 | `Box(243)` | `Box(39)` |
28
+
29
+ 각 데이터셋은 `<환경>/<tier>/data/main_data.hdf5`와 `metadata.json`으로
30
+ 구성됩니다. 관측 센서는 의미가 섞이지 않도록 별도의 `Tuple` leaf로 보존하며,
31
+ 센서가 하나인 환경은 단일 `Box`를 사용합니다.
32
+
33
+ SoccerTwos와 DungeonEscape는 공유 정책을 사용하는 분산형 다중 에이전트
34
+ 환경입니다. 각 물리 에이전트는 ego episode 하나로 저장되며, 같은 경기나 협동
35
+ 그룹은 함께 제공되는 match/group metadata로 연결됩니다.
36
+
37
+ DungeonEscape·SoccerTwos·Pyramids·Crawler·PushBlock·3DBallHard·Worm·Walker는
38
+ `expert` 외에 스톡 정책을 점진적으로 열화시킨 `medium`/`simple` 품질 래더
39
+ 티어도 제공합니다. 티어는 Minari 정규화 스킬 기준(`simple` 0.60, `medium` 0.80,
40
+ `expert` = 1.0)으로 보정됩니다. 열화 방식은 씬에 따라 다릅니다(seed 2310000) —
41
+ 연속 제어 씬은 에피소드마다 보정된 가우시안 노이즈 범위를 씁니다. 단일 에이전트
42
+ 목표 게임인 PushBlock과 Pyramids는 정책 자체 로짓에 대한 softmax temperature
43
+ `T`(`softmax(logits/T)`)를 보정된 범위에서 **에피소드마다** 샘플링(각 에피소드가
44
+ 자체 스킬 레벨)합니다. 협동 DungeonEscape은 경기마다 보정된
45
+ 범위에서 softmax temperature `T`를 **그룹 단위로** 샘플링해(그룹의 세 에이전트가
46
+ 같은 `T`를 공유 — 한 시점의 초기 체크포인트처럼 하나의 일관된 스킬 레벨) 역시
47
+ step-reward로 채점하며, 자기대전 SoccerTwos는 매치마다 팀별로 softmax temperature
48
+ `T`를 (더 넓은) 보정 범위에서 **독립 샘플링**합니다(두 팀이 독립적으로 뽑아 매치마다
49
+ 실력차가 생김 — 여러 체크포인트가 붙는 리그전처럼). 목표
50
+ 게임은 에피소드 리턴이 이봉(해결 vs 타임아웃)이라 **step-reward**(리턴/길이)로
51
+ 채점합니다. 각 티어 평균은 목표에 맞으면서 에피소드들이 스킬 스펙트럼을 연속적으로
52
+ 덮습니다. 티어별 수치는 영문 카드(README.md)의 per-environment quality ladder
53
+ 표를 참고하세요.
54
+
55
+ ## 동반 저장소
56
+
57
+ - 모델 제거 Unity 빌드와 스톡 ONNX:
58
+ [ccnets/causal-gpt-rl-unity-envs](https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-envs)
59
+ - 학습된 Causal GPT-RL 정책:
60
+ [ccnets/causal-gpt-rl-unity](https://huggingface.co/ccnets/causal-gpt-rl-unity)
61
+ - 수집 및 Minari 패키징 코드:
62
+ [examples/unity_collection](https://github.com/ccnets-team/causal-gpt-rl/tree/main/examples/unity_collection)
63
+
64
+ ## 로드 예시
65
+
66
+ ```python
67
+ from pathlib import Path
68
+ from huggingface_hub import snapshot_download
69
+ import minari
70
+
71
+ root = Path.home() / ".minari" / "datasets" / "unity"
72
+ snapshot_download(
73
+ repo_id="ccnets/causal-gpt-rl-unity-datasets",
74
+ repo_type="dataset",
75
+ allow_patterns="crawler/**",
76
+ local_dir=root,
77
+ )
78
+ dataset = minari.load_dataset("unity/crawler/expert-v0")
79
+ print(dataset.total_episodes, dataset.total_steps)
80
+ ```
81
+
82
+ ## 라이선스
83
+
84
+ 궤적 데이터는 Creative Commons Attribution 4.0 International
85
+ (CC-BY-4.0)로 배포됩니다. *ccnets — Causal GPT-RL*과 Unity ML-Agents 원본
86
+ 환경을 표기해 주세요.