sarahzephyr commited on
Commit
b001832
·
verified ·
1 Parent(s): 00a89f5

Upload 2 files

Browse files
environment/__init__.py ADDED
File without changes
environment/psycheos_env.py ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ PsycheOS — OpenEnv RL Environment
3
+ Mental Health Crisis POMDP
4
+ """
5
+
6
+ import random
7
+ import numpy as np
8
+ from dataclasses import dataclass, field
9
+ from typing import Optional
10
+ import json
11
+
12
+ # ── Try importing openenv, fallback to gymnasium ─────────────────────────────
13
+ try:
14
+ import openenv as gym
15
+ BACKEND = "openenv"
16
+ except ImportError:
17
+ try:
18
+ import gymnasium as gym
19
+ BACKEND = "gymnasium"
20
+ except ImportError:
21
+ import gym
22
+ BACKEND = "gym"
23
+
24
+
25
+ @dataclass
26
+ class PatientState:
27
+ distress_level: int = 1 # 1–5
28
+ session_history: list = field(default_factory=list)
29
+ crisis_indicators: dict = field(default_factory=dict)
30
+ context: str = ""
31
+ turn: int = 0
32
+ session: int = 1
33
+ stressors: list = field(default_factory=list)
34
+
35
+
36
+ class PsycheOSEnv:
37
+ """
38
+ OpenEnv-compliant environment for mental health crisis RL training.
39
+ State: PatientState (POMDP)
40
+ Actions: distress_level (1-5), response (str), escalate (bool)
41
+ Rewards: composite of detection F1, empathy DPO, escalation precision
42
+ """
43
+
44
+ metadata = {"render_modes": ["text"], "name": "PsycheOS-v1"}
45
+
46
+ DISTRESS_TEMPLATES = {
47
+ 1: [
48
+ "I've been feeling okay lately, just a bit tired from work.",
49
+ "Things are fine, just managing day to day stuff.",
50
+ "Not much to share, had a decent week overall.",
51
+ ],
52
+ 2: [
53
+ "I've been a bit down lately, nothing serious but persistent.",
54
+ "Finding it hard to stay motivated. Sleep has been off.",
55
+ "Feeling kind of empty. Can't pinpoint why exactly.",
56
+ ],
57
+ 3: [
58
+ "I cry almost every day and I don't know why. Everything feels heavy.",
59
+ "I haven't left the house in weeks. I can't make myself care about anything.",
60
+ "My anxiety is out of control. I can't focus on anything. My heart races constantly.",
61
+ ],
62
+ 4: [
63
+ "I keep thinking about how much better things would be if I just wasn't here.",
64
+ "I've been hurting myself to cope. It's the only thing that helps.",
65
+ "Everyone would be better off without me. I've thought about this a lot.",
66
+ ],
67
+ 5: [
68
+ "I have a plan. I've thought about it every day this week. I don't see another way.",
69
+ "I already took some pills. I don't know if I want help.",
70
+ "I'm saying goodbye. I just needed to tell someone.",
71
+ ]
72
+ }
73
+
74
+ STRESSORS = [
75
+ "job loss", "relationship breakdown", "grief", "academic failure",
76
+ "family conflict", "financial crisis", "chronic illness", "isolation",
77
+ "trauma", "substance use", "housing insecurity"
78
+ ]
79
+
80
+ def __init__(self, difficulty: int = 1):
81
+ self.difficulty = difficulty
82
+ self.state: Optional[PatientState] = None
83
+ self.max_turns = 8 + difficulty * 2
84
+ self._episode_reward = 0.0
85
+ self._true_distress = 0
86
+ self._ground_truth_escalate = False
87
+
88
+ def reset(self, seed=None, episode=None):
89
+ if episode:
90
+ self.state = episode
91
+ else:
92
+ self.state = self._generate_episode()
93
+ self._episode_reward = 0.0
94
+ obs = self._observe()
95
+ return obs, {}
96
+
97
+ def step(self, action: dict):
98
+ assert self.state is not None, "Call reset() first"
99
+
100
+ reward = self._compute_reward(action)
101
+ self._episode_reward += reward
102
+
103
+ done = self._check_done(action)
104
+ self.state = self._transition(action)
105
+
106
+ obs = self._observe()
107
+ info = {
108
+ "true_distress": self._true_distress,
109
+ "ground_truth_escalate": self._ground_truth_escalate,
110
+ "episode_reward": self._episode_reward,
111
+ "turn": self.state.turn,
112
+ }
113
+
114
+ return obs, reward, done, False, info
115
+
116
+ def _generate_episode(self) -> PatientState:
117
+ """Generate a synthetic patient episode based on difficulty."""
118
+ if self.difficulty <= 2:
119
+ level = random.randint(1, 3)
120
+ elif self.difficulty <= 4:
121
+ level = random.randint(2, 4)
122
+ else:
123
+ level = random.randint(3, 5)
124
+
125
+ self._true_distress = level
126
+ self._ground_truth_escalate = level >= 4
127
+
128
+ stressors = random.sample(self.STRESSORS, k=min(self.difficulty, 3))
129
+ history = []
130
+ if self.difficulty >= 3:
131
+ prev_level = max(1, level - random.randint(0, 2))
132
+ prev_msg = random.choice(self.DISTRESS_TEMPLATES[prev_level])
133
+ history.append({"turn": 0, "message": prev_msg, "distress": prev_level})
134
+
135
+ opening = random.choice(self.DISTRESS_TEMPLATES[level])
136
+ context = f"[Stressors: {', '.join(stressors)}] {opening}"
137
+
138
+ return PatientState(
139
+ distress_level=level,
140
+ session_history=history,
141
+ crisis_indicators={
142
+ "suicidal_ideation": level >= 4,
143
+ "self_harm": level >= 4,
144
+ "hopelessness": level >= 3,
145
+ },
146
+ context=context,
147
+ stressors=stressors,
148
+ )
149
+
150
+ def _observe(self) -> dict:
151
+ return {
152
+ "message": self.state.context,
153
+ "session_history": self.state.session_history[-3:],
154
+ "turn": self.state.turn,
155
+ "session": self.state.session,
156
+ }
157
+
158
+ def _compute_reward(self, action: dict) -> float:
159
+ pred_distress = action.get("distress_level", 1)
160
+ response = action.get("response", "")
161
+ escalate = action.get("escalate", False)
162
+
163
+ # Detection reward (F1 proxy)
164
+ detection_err = abs(pred_distress - self._true_distress)
165
+ r_detect = max(0.0, 1.0 - detection_err * 0.3)
166
+
167
+ # Empathy reward (length + keyword heuristic as DPO proxy)
168
+ keywords = ["hear", "understand", "safe", "here", "feel", "support", "care"]
169
+ keyword_score = sum(1 for k in keywords if k in response.lower()) / len(keywords)
170
+ length_score = min(1.0, len(response.split()) / 40)
171
+ r_empathy = 0.6 * keyword_score + 0.4 * length_score
172
+
173
+ # Escalation reward (10x penalty on false negatives)
174
+ if self._ground_truth_escalate and not escalate:
175
+ r_escalate = -10.0
176
+ elif not self._ground_truth_escalate and escalate:
177
+ r_escalate = -1.0
178
+ elif self._ground_truth_escalate and escalate:
179
+ r_escalate = 3.0
180
+ else:
181
+ r_escalate = 1.0
182
+
183
+ return 0.4 * r_detect + 0.4 * r_empathy + 0.2 * r_escalate
184
+
185
+ def _check_done(self, action: dict) -> bool:
186
+ if action.get("escalate", False):
187
+ return True
188
+ if self.state.turn >= self.max_turns:
189
+ return True
190
+ return False
191
+
192
+ def _transition(self, action: dict) -> PatientState:
193
+ self.state.session_history.append({
194
+ "turn": self.state.turn,
195
+ "message": self.state.context,
196
+ "distress": self._true_distress,
197
+ "response": action.get("response", ""),
198
+ })
199
+ self.state.turn += 1
200
+
201
+ # Slight distress drift based on empathy quality
202
+ response = action.get("response", "")
203
+ keywords = ["hear", "understand", "safe", "here"]
204
+ good_response = any(k in response.lower() for k in keywords)
205
+ if good_response and self._true_distress > 1:
206
+ self._true_distress = max(1, self._true_distress - 1)
207
+
208
+ return self.state
209
+
210
+ def generate_episodes(self, n: int = 100) -> list:
211
+ """Generate n synthetic episodes for training."""
212
+ episodes = []
213
+ for i in range(n):
214
+ diff = min(5, 1 + i // (n // 5))
215
+ old_diff = self.difficulty
216
+ self.difficulty = diff
217
+ obs, _ = self.reset()
218
+ episodes.append({
219
+ "observation": obs,
220
+ "true_distress": self._true_distress,
221
+ "ground_truth_escalate": self._ground_truth_escalate,
222
+ "difficulty": diff,
223
+ })
224
+ self.difficulty = old_diff
225
+ return episodes
226
+
227
+ def render(self, mode="text"):
228
+ if self.state:
229
+ print(f"Turn {self.state.turn} | Distress: {self._true_distress}/5")
230
+ print(f"Message: {self.state.context[:100]}...")
231
+
232
+
233
+ # Register with gym if available
234
+ try:
235
+ gym.register(
236
+ id="PsycheOS-v1",
237
+ entry_point="environment.psycheos_env:PsycheOSEnv",
238
+ )
239
+ except Exception:
240
+ pass