Run formatter, bump version

Author: dbobrenko

prompting/llms/vllm_llm.py

@@ -138,9 +138,7 @@ async def generate_logits(
             return {}
 
         logprobs = outputs[0].outputs[0].logprobs[0]
-        token_logprobs = {
-            self.tokenizer.decode([token]): logprob.logprob for token, logprob in logprobs.items()
-        }
+        token_logprobs = {self.tokenizer.decode([token]): logprob.logprob for token, logprob in logprobs.items()}
         sorted_token_logprobs = dict(sorted(token_logprobs.items(), key=lambda item: item[1], reverse=True))
         return sorted_token_logprobs, prompt
 

prompting/rewards/exact_match.py

@@ -1,10 +1,10 @@
 import random
 
 import numpy as np
-from loguru import logger
-from openai.types.chat import ChatCompletionChunk
 import torch
 import torch.nn.functional as F
+from loguru import logger
+from openai.types.chat import ChatCompletionChunk
 
 from prompting.llms.model_manager import ModelManager
 from prompting.rewards.reward import BaseRewardModel, BatchRewardOutput
@@ -18,7 +18,7 @@
 INCORRECT_PENALTY = -2
 MIN_SMOOTH_PENALTY_SCALE = 0.6
 MIN_TIME_PENALTY_SCALE = 0.3
-VERIFICATION_THRESH_CONTAINS = 0.90
+VERIFICATION_THRESH_CONTAINS = 0.92
 VERIFICATION_THRESH_SIM = 0.90
 MIN_VERIFY_TOKENS = 10
 MAX_VERIFY_TOKENS = 30
@@ -177,7 +177,7 @@ async def reward(
         )
         logger.debug(f"Logits rewards: {reward_output.model_dump()}")
         return reward_output
-    
+
     @staticmethod
     def sample_verification_indices(completion_length: int) -> list[int]:
         """Sample random indices for verification, always add eos_token index."""
@@ -245,7 +245,7 @@ def verify_logit_similarity(
         """Similarity between candidate and ground-truth logprobs."""
         if not gt_logits:
             return 0.0
-        
+
         if len(candidate_logits) != len(gt_logits):
             return 0.0
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "prompting"
-version = "2.19.0"
+version = "2.19.1"
 description = "Subnetwork 1 runs on Bittensor and is maintained by Macrocosmos. It's an effort to create decentralised AI"
 authors = ["Kalei Brady, Dmytro Bobrenko, Felix Quinque, Steffen Cruz, Richard Wardle"]
 readme = "README.md"

tests/prompting/rewards/test_exact_match.py

@@ -301,4 +301,4 @@ def test_sample_verification_indices_properties(completion_length):
     sample_indices = indices[:-1]
     assert all(0 <= idx < completion_length for idx in sample_indices)
     # No duplicates overall.
-    assert len(set(indices)) == len(indices)
+    assert len(set(indices)) == len(indices)