Merge pull request #85 from opentensor/staging

2.1.5 Release

neurons/validators/validator.py

@@ -194,7 +194,7 @@ def __init__(self):
                 MockRewardModel(RewardModelType.nsfw.value),
             ]
             self.penalty_functions = [
-                TaskValidationPenaltyModel(max_penalty=0.6),
+                TaskValidationPenaltyModel(max_penalty=0.75),
                 ContentMatchPenaltyModel(max_penalty=0.2),
                 KeywordMatchPenaltyModel(max_penalty=1),
             ]

prompting/validators/__init__.py

@@ -27,7 +27,7 @@
 from . import event
 from . import dataset
 
-__version__ = "2.1.4"
+__version__ = "2.1.5"
 version_split = __version__.split(".")
 __spec_version__ = (
     (1000 * int(version_split[0]))

prompting/validators/criteria.py

@@ -241,4 +241,4 @@ def evaluate(self, completions: List[str]) -> torch.FloatTensor:
         return penalties
 
     def compose_text(self) -> str:
-        return self.text.format(layout_type=self.layout_type)
+        return self.text.format(layout_type=self.layout_type.value)

prompting/validators/forward.py

@@ -221,7 +221,7 @@ async def run_step(self, task: Task, k: int, timeout: float, exclude: list = [])
     return event
 
 
-async def forward(self):
+async def questions_and_answers_around_summary_flow(self):
     # Obtain a unique context from the dataset.
     data = next(self.dataset)["text"]
 
@@ -272,3 +272,8 @@ async def forward(self):
         )
 
         exclude += qa_event["uids"]
+
+
+async def forward(self):
+    # Definition of flow to be executed at forward step
+    await questions_and_answers_around_summary_flow(self)

prompting/validators/reward/blacklist.py

@@ -43,9 +43,9 @@ def name(self) -> str:
 
     def __init__(
         self,
-        boundary: float = 6,
+        boundary: float = 40,
         n_min: int = 5,
-        n_max: int = 14,
+        n_max: int = 10,
         word_limit: int = 2000,
         A: float = 1.3,
         preprocess: str = "[^(\\w|\\s)]",
@@ -213,7 +213,7 @@ def calculate_significance(self) -> dict:
                 if len(decoded_ngram.split()) >= self.n_min:
                     # calculate significance score for ngram
                     significance_scores[decoded_ngram] = (
-                        self.A ** (len(decoded_ngram) - 1)
+                        self.A ** (len(decoded_ngram.split()) - 1)
                         * ((count[0] + count[1]) / self.num_completion)
                         * self.frequency_multiplier
                     )
@@ -302,7 +302,7 @@ def reward(self, prompt: str, completion: str, name: str) -> BlacklistRewardEven
                 and fuzz.partial_ratio(ngram, completion.lower())
                 > self.partial_ratio_boundary
             ):
-                reward_event.reward = 1
+                reward_event.reward = 0
                 reward_event.matched_ngram = ngram
                 reward_event.significance_score = score
                 return reward_event

prompting/validators/reward/reward.py

@@ -178,7 +178,7 @@ def apply(
             bt.logging.warning(
                 f"The tensor from {self.name} contains NaN values: {filled_rewards_normalized}"
             )
-            filled_rewards_normalized.nan_to_num_(nan=0.0)
+            filled_rewards_normalized = filled_rewards_normalized.nan_to_num_(nan=0.0)
 
         # Return the filled rewards.
         return filled_rewards_normalized, reward_events

prompting/validators/tasks.py

@@ -28,6 +28,8 @@
     ContentMatchTypeEnum,
     SimpleResponseLayoutCriteria,
     MatchContentCriteria,
+    MatchLayoutCriteria,
+    LayoutMatchTypeEnum,
 )
 
 
@@ -122,24 +124,61 @@ def compose_prompt(self) -> str:
 
 
 def create_summarization_task(base_text: str) -> SummaryTask:
-    possible_criterias = [
-        MatchLengthCriteria(
-            penalty=0.1,
-            target_length=random.randint(50, 200),
-            unit=TextLengthUnitEnum.WORDS,
-        ),
-        MatchLengthCriteria(
-            penalty=0.1,
-            target_length=random.randint(4, 8),
-            unit=TextLengthUnitEnum.SENTENCES,
-        ),
-    ]
+    # scope 1: bullet points, scope 2: numbered list, scope 3: simple layout
+    scope = random.randint(1, 3)
+
+    select_bullet_point_layout = scope == 1
+    select_numbered_list_layout = scope == 2
+
+    # scope 1 or 2: define criteria set for bullet points or numbered list
+    if select_bullet_point_layout or select_numbered_list_layout:
+        if select_bullet_point_layout:
+            layout_criteria = MatchLayoutCriteria(
+                layout_type=LayoutMatchTypeEnum.UNORDERED_LIST,
+                penalty=0.5,
+                text="Your response should be ordered in format of bullet points.",
+            )
+        else:
+            layout_criteria = MatchLayoutCriteria(
+                layout_type=LayoutMatchTypeEnum.NUMBERED_LIST,
+                penalty=0.5,
+            )
+
+        possible_other_criterion = [
+            MatchLengthCriteria(
+                penalty=0.25,
+                target_length=random.randint(100, 200),
+                unit=TextLengthUnitEnum.WORDS,
+            ),
+            MatchLengthCriteria(
+                penalty=0.25,
+                target_length=random.randint(8, 12),
+                unit=TextLengthUnitEnum.SENTENCES,
+            ),
+        ]
+    # scope 3: define criteria set for simple layout
+    else:
+        layout_criteria = SimpleResponseLayoutCriteria(penalty=0.5)
+
+        possible_other_criterion = [
+            MatchLengthCriteria(
+                penalty=0.25,
+                target_length=random.randint(50, 200),
+                unit=TextLengthUnitEnum.WORDS,
+            ),
+            MatchLengthCriteria(
+                penalty=0.25,
+                target_length=random.randint(4, 8),
+                unit=TextLengthUnitEnum.SENTENCES,
+            ),
+        ]
 
-    sampled_criterias = random.sample(possible_criterias, 1)
+    random_sampled_criterion = random.sample(possible_other_criterion, 1)
+    defined_criteria = [layout_criteria] + random_sampled_criterion
 
     return SummaryTask(
         base_text=base_text,
-        criteria=sampled_criterias,
+        criteria=defined_criteria,
         task_type="summarization",
         task_name="augment",
     )
@@ -192,12 +231,12 @@ def create_qg_task(base_text: str, index: int) -> QuestionGenerationTask:
 
     other_random_criteria = [
         MatchLengthCriteria(
-            penalty=0.1,
+            penalty=0.25,
             target_length=random.randint(10, 40),
             unit=TextLengthUnitEnum.WORDS,
         ),
         MatchLengthCriteria(
-            penalty=0.1,
+            penalty=0.25,
             target_length=random.randint(40, 150),
             unit=TextLengthUnitEnum.CHARACTERS,
         ),
@@ -221,16 +260,16 @@ def create_qa_task(base_text: str, index: int) -> QuestionAnswerTask:
     answer_should_not_include_criteria = MatchContentCriteria(
         words_array=["?"],
         n_words=1,
-        penalty=0.2,
+        penalty=0.25,
         contentMatchType=ContentMatchTypeEnum.INCLUDES,
         negate_match=True,
         text="Your response should not include any question marks",
     )
 
-    simple_response_layout_criteria = SimpleResponseLayoutCriteria(penalty=0.2)
+    simple_response_layout_criteria = SimpleResponseLayoutCriteria(penalty=0.25)
 
     words_criteria = MatchLengthCriteria(
-        penalty=0.2,
+        penalty=0.25,
         target_length=random.randint(50, 200),
         unit=TextLengthUnitEnum.WORDS,
     )