graceful SentenceTransformer fallback on HF Spaces

envs/arbitragent_env.py

@@ -45,7 +45,10 @@ class ArbitrAgentEnv(Env):
 
     def __init__(self, data_path: str = "training/data/selfplay_states.json", seed=None):
         self.data_path = data_path
-        self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        try:
+            self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        except Exception:
+            self.encoder = None
         if seed is not None:
             random.seed(seed)
             np.random.seed(seed)
@@ -104,6 +107,8 @@ class ArbitrAgentEnv(Env):
 
     def _accuracy_reward(self, action: str) -> float:
         """Cosine similarity between action embedding and human action embedding."""
+        if self.encoder is None:
+            return 0.0
         state_text = self.current_state.get("state_text", "")
         human_action_text = _extract_human_orders(state_text)
         action_emb = self.encoder.encode(action, convert_to_numpy=True)
@@ -204,6 +209,8 @@ Your task: Propose a move. If you detect a bluff, use coalition pressure; otherw
 
     def _get_observation(self):
         text = self._get_state_text()
+        if self.encoder is None:
+            return np.zeros(384, dtype=np.float32)
         emb = self.encoder.encode(text, convert_to_numpy=True)
         return emb.astype(np.float32)
 

envs/contractor_env.py

@@ -24,7 +24,10 @@ class ContractorNegotiationEnv(Env):
     def __init__(self, n_contractors=5, budget=10000, seed=None):
         self.n_contractors = n_contractors
         self.budget = budget
-        self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        try:
+            self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        except Exception:
+            self.encoder = None
         if seed:
             random.seed(seed)
             np.random.seed(seed)
@@ -145,6 +148,8 @@ class ContractorNegotiationEnv(Env):
 
     def _get_observation(self):
         text = self._get_state_text()
+        if self.encoder is None:
+            return np.zeros(384, dtype=np.float32)
         emb = self.encoder.encode(text, convert_to_numpy=True)
         return emb.astype(np.float32)
 

envs/diplomacy_env.py

@@ -19,7 +19,10 @@ class DiplomacyNegotiationEnv(Env):
     def __init__(self, power_name: str = "ENGLAND", seed: int | None = None):
         self._reset_random_power = power_name.upper() == "ENGLAND"  # default: vary power on reset for non-hardcoded obs
         self.power_name = power_name.upper()
-        self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        try:
+            self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        except Exception:
+            self.encoder = None
         self.game: Game | None = None
         self.current_phase: int = 0
         self.prev_sc_count: int = 0
@@ -149,6 +152,8 @@ class DiplomacyNegotiationEnv(Env):
     def _get_observation(self) -> np.ndarray:
         """Return a 384-dim MiniLM embedding of the current game state text."""
         text = self._get_state_text()
+        if self.encoder is None:
+            return np.zeros(384, dtype=np.float32)
         embedding = self.encoder.encode(text, convert_to_numpy=True)
         # Ensure consistent dtype for downstream RL code.
         return embedding.astype(np.float32)

envs/human_imitation_env.py

@@ -17,7 +17,10 @@ from sentence_transformers import SentenceTransformer
 class HumanImitationEnv(Env):
     def __init__(self, data_path="training/data/selfplay_states.json", seed=None):
         self.data_path = data_path
-        self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        try:
+            self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        except Exception:
+            self.encoder = None
         if seed is not None:
             random.seed(seed)
             np.random.seed(seed)
@@ -115,6 +118,8 @@ Explain your reasoning and state your intended orders."""
 
     def _get_observation(self):
         text = self._get_state_text()
+        if self.encoder is None:
+            return np.zeros(384, dtype=np.float32)
         emb = self.encoder.encode(text, convert_to_numpy=True)
         return emb.astype(np.float32)
 
@@ -156,7 +161,10 @@ from sentence_transformers import SentenceTransformer
 class HumanImitationEnv(Env):
     def __init__(self, data_path="training/data/selfplay_states.json", seed=None):
         self.data_path = data_path
-        self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        try:
+            self.encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        except Exception:
+            self.encoder = None
         if seed is not None:
             random.seed(seed)
             np.random.seed(seed)
@@ -254,6 +262,8 @@ Explain your reasoning and state your intended orders."""
 
     def _get_observation(self):
         text = self._get_state_text()
+        if self.encoder is None:
+            return np.zeros(384, dtype=np.float32)
         emb = self.encoder.encode(text, convert_to_numpy=True)
         return emb.astype(np.float32)