sarahzephyr commited on
Commit
d00a723
·
verified ·
1 Parent(s): 1f13f05

Upload openenv.yaml

Browse files
Files changed (1) hide show
  1. openenv.yaml +128 -0
openenv.yaml ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: GodMode
3
+ emoji: 🏛️
4
+ colorFrom: purple
5
+ colorTo: pink
6
+ sdk: docker
7
+ pinned: false
8
+ license: mit
9
+
10
+ app_port: 7860
11
+
12
+ tags:
13
+ - civilization
14
+ - multi-agent
15
+ - long-horizon
16
+ - strategy
17
+ - reinforcement-learning
18
+ - grpo
19
+
20
+ short_description: >
21
+ A civilization management RL environment where an LLM learns to rule
22
+ a nation across 30 turns with strategy, diplomacy, trade, and war.
23
+ ---
24
+
25
+ name: GodMode
26
+ version: "1.0.0"
27
+ description: >
28
+ A civilization management RL environment where an LLM learns to rule
29
+ a nation across 30 turns. Manages food, gold, military, diplomacy,
30
+ trade, and survives disasters. Four adversarial rival civilizations
31
+ with hidden intentions. Long-horizon planning with delayed consequences.
32
+
33
+ author: Mesa Sarah Vasantha Zephyr
34
+ license: MIT
35
+
36
+ environment:
37
+ id: GodMode-v1
38
+ entry_point: environment.godmode_env:GodModeEnv
39
+ max_episode_steps: 30
40
+ server: server/app.py
41
+
42
+ observation_space:
43
+ type: dict
44
+ fields:
45
+ turn:
46
+ type: int
47
+ season:
48
+ type: text
49
+ active_disaster:
50
+ type: text
51
+ player:
52
+ type: dict
53
+ description: "food, gold, military, population, tech, happiness, territory, relations"
54
+ rivals:
55
+ type: list
56
+ description: "4 rival civilizations with approximate stats only; hidden intent not visible"
57
+ recent_events:
58
+ type: list
59
+ available_actions:
60
+ type: list
61
+
62
+ action_space:
63
+ type: dict
64
+ fields:
65
+ reasoning:
66
+ type: text
67
+ description: "Chain-of-thought before acting"
68
+ action:
69
+ type: discrete
70
+ options:
71
+ - farm
72
+ - mine
73
+ - recruit
74
+ - research
75
+ - trade
76
+ - diplomacy
77
+ - spy
78
+ - attack
79
+ - defend
80
+ - build
81
+ target:
82
+ type: text
83
+ description: "Target rival civilization name"
84
+ investment:
85
+ type: float
86
+ description: "Gold to invest (0–50)"
87
+
88
+ reward:
89
+ type: composite
90
+ components:
91
+ - name: survival
92
+ weight: 0.30
93
+ description: "Stay alive — population and food above threshold"
94
+ - name: growth
95
+ weight: 0.25
96
+ description: "Expand population, territory, technology, and gold"
97
+ - name: happiness
98
+ weight: 0.15
99
+ description: "Keep citizens happy and prevent rebellion"
100
+ - name: diplomacy
101
+ weight: 0.15
102
+ description: "Alliances rewarded, wars penalized"
103
+ - name: longevity
104
+ weight: 0.10
105
+ description: "Bonus for surviving later turns"
106
+ - name: diversity
107
+ weight: 0.05
108
+ description: "Penalizes repetitive strategies"
109
+
110
+ training:
111
+ framework: GRPO
112
+ warmup: "SFT (50 steps on expert traces)"
113
+ library: "HuggingFace TRL + Unsloth"
114
+ base_model: "unsloth/llama-3-8b-instruct"
115
+ adapter: "LoRA (r=16, alpha=32)"
116
+ quantization: "4-bit"
117
+ save_method: "save_pretrained_merged"
118
+ script: training/grpo_train.py
119
+ notebook: training/colab_training.ipynb
120
+
121
+ themes:
122
+ - "Theme 1 - Multi-Agent"
123
+ - "Theme 2 - Long-Horizon Planning"
124
+ - "Theme 3 - World Modeling"
125
+ - "Theme 4 - Self-Improvement"
126
+ - "Theme 5 - Wild Card"
127
+
128
+ space_url: "https://huggingface.co/spaces/mesazephyr/godmode"