Remerge main into pre-staging

prompting/rewards/ordinal.py

@@ -7,22 +7,16 @@
 class OrdinalRewardModel(BaseRewardModel):
     @property
     def name(self) -> str:
-        return "category_distance"
+        return "ordinal"
 
     def __init__(self, **kwargs):
         super().__init__()
         #TODO: Expand to allow for more than 3 classes (Must also adjust dataset/review.py)
-        self.sentiments = [
-            "casual",
-            "basic",
-            "silly",
-            "random",
-            "thoughtful",
-            "serious",
-            "rushed",
+        self.sentiments = [ 
+            "positive",
+            "neutral",
+            "negative",
         ]
-        #NOTE: These sentimens are not the same as the sentiments defined in the dataset/review.py file. These are the subtopic
-
 
     def reward(self, reference: str, completions: List[str]) -> BatchRewardOutput:
         """Compute difference scores given a completion and reference pair."""
@@ -31,10 +25,11 @@ def reward(self, reference: str, completions: List[str]) -> BatchRewardOutput:
         classes = self.sentiments
         for completion in completions:
             t0 = time.time()
-
+            completion = completion.lower()
             # Check if exactly one answer can be found in the completion
             if sum(option in completion for option in classes) == 1:
-                reward = abs(classes.index(reference) - classes.index(completion))
+                answer = [option for option in classes if option in completion][0]
+                reward = 1-abs(classes.index(reference) - classes.index(answer))/(len(classes)-1)
             else:
                 reward = 0 
             timings.append(time.time() - t0)
@@ -44,7 +39,7 @@ def reward(self, reference: str, completions: List[str]) -> BatchRewardOutput:
             rewards=torch.FloatTensor(rewards),
             timings=torch.FloatTensor(timings),
             extra_info={
-                "type": "math",
+                "type": "ordinal",
             },
         )
-        return output
+        return output

prompting/rewards/reward.py

@@ -100,10 +100,7 @@ def total_reward(self) -> torch.FloatTensor:
         """Combines the rewards from all the reward models into a single reward tensor"""
 
         # TODO: How would using the Agent as a reward model fit into this flow?
-        # Compute the rewards for the responses given the prompt
-        rewards = torch.zeros_like(
-            self.response_event.uids, dtype=torch.float32, device=self.device
-        )
+        # Compute the rewards for the responses given the prompt        
         rewards = torch.zeros_like(
             self.response_event.uids, dtype=torch.float32, device=self.device
         )

prompting/tasks/challenge_templates/base.py

@@ -11,7 +11,7 @@ def next(self, query: str):
         self.fields["query"] = [query]
         return self.get_template().format(
             **{field: random.choice(entries) for field, entries in self.fields.items()}
-        )
+        ).split("<end>")[0]
 
     def get_template(self):
         return random.choice(self.templates)

prompting/tasks/challenge_templates/math.py

@@ -6,10 +6,10 @@ class MathChallengeTemplate(ChallengeTemplate):
     def __init__(self):
         super().__init__()
         self.templates = [
-            "{greeting}{greeting_punctuation}{query}{request}<end>",
-            "{query}<end>{greeting}{greeting_punctuation}{request}",
-            "{greeting}{greeting_punctuation}{query}<end>{request}",
-            "{query}{request}<end>{greeting}{greeting_punctuation}",
+            "{greeting}{greeting_punctuation}{query}{whitespace}{request}<end>",
+            "{query}<end>{greeting}{greeting_punctuation}{request}{whitespace}",
+            "{greeting}{greeting_punctuation}{query}<end>{request}{whitespace}",
+            "{query}{whitespace}{request}<end>{greeting}{greeting_punctuation}",
         ]
         self.fields = {
             "greeting": [
@@ -181,4 +181,11 @@ def __init__(self):
                 "Can you help me understand this, please?",
                 "I could use your help figuring this out.",
             ],
+            "whitespace": [
+                "",
+                " ",
+                " ",
+                "\n",
+                "\n\n",
+            ]
         }

prompting/tasks/challenge_templates/sentiment.py

@@ -6,8 +6,8 @@ class SentimentChallengeTemplate(ChallengeTemplate):
     def __init__(self):
         super().__init__()
         self.templates = [
-            "{greeting}{greeting_punctuation}{request}{query}<end>",
-            "{request}{query}<end>{greeting}{greeting_punctuation}",
+            "{greeting}{greeting_punctuation}{request}{whitespace}{query}<end>",
+            "{request}{whitespace}{query}<end>{greeting}{greeting_punctuation}",
         ]
         self.fields = {
             "greeting": [
@@ -176,4 +176,11 @@ def __init__(self):
                 "Can you evaluate the sentiment of this text?",
                 "I'm interested in the sentiment analysis of this text.",
             ],
+            "whitespace": [
+                "",
+                " ",
+                " ",
+                "\n",
+                "\n\n",
+            ]
         }

prompting/tasks/sentiment.py

@@ -12,7 +12,6 @@ class SentimentAnalysisTask(Task):
     desc = "get help analyzing the sentiment of a review"
     goal = "to get the sentiment to the following review"
     challenge_type = 'paraphrase'
-    challenge_type = 'paraphrase'
     challenge_template = SentimentChallengeTemplate()
 
     reward_definition = [

prompting/tasks/translate.py

@@ -71,6 +71,9 @@ def translate_to_random_language(self, content: str, from_code:str='en') -> Tupl
 
 @dataclass
 class TranslationTask(Task):
+    challenge_type = 'query'
+    static_reference = True
+    static_query = True
     name = "translation"
     desc = "get translation help"
     goal = "to get the translation for the given piece of text"