Merge pull request #7 from prabha-git/mixtral

Mixtral

GoogleEvaluator.py

@@ -56,8 +56,8 @@ async def get_response_from_model(self, prompt):
 if __name__ == "__main__":
     # Tons of defaults set, check out the LLMNeedleHaystackTester's init for more info
     ht = GoogleEvaluator(model_name='gemini-pro', evaluation_method='substring_match',
-                         save_results_dir='gemini-pro-results-run2',
-                         save_contexts_dir='gemini-pro-context-run2'
+                         save_results_dir='results/gemini-pro/gemini-pro-results-run2',
+                         save_contexts_dir='contexts/gemini-pro/gemini-pro-context-run2'
                          )
 
     ht.start_test()

LLMNeedleHaystackTester.py

@@ -219,16 +219,17 @@ async def evaluate_and_log(self, context_length, depth_percent):
             if not os.path.exists(self.save_contexts_dir):
                 os.makedirs(self.save_contexts_dir)
 
-            with open(f'{self.save_contexts_dir}/{self.save_contexts_dir}_{context_file_location}_context.txt', 'w') as f:
+            with open(f'{self.save_contexts_dir}/{context_file_location}_context.txt', 'w') as f:
                 f.write(context)
 
+
         if self.save_results:
             # Save the context to file for retesting
             if not os.path.exists(self.save_results_dir):
                 os.makedirs(self.save_results_dir)
 
             # Save the result to file for retesting
-            with open(f'{self.save_results_dir}/{self.save_results_dir}_{context_file_location}_results.json', 'w') as f:
+            with open(f'{self.save_results_dir}/{context_file_location}_results.json', 'w') as f:
                 json.dump(results, f)
 
         if self.seconds_to_sleep_between_completions:

OSS_TogetherEvaluator.py

@@ -0,0 +1,62 @@
+import os
+import tiktoken
+from LLMNeedleHaystackTester import LLMNeedleHaystackTester
+from langchain_together import Together
+import time
+
+
+
+class OSS_TogetherEvaluator(LLMNeedleHaystackTester):
+    def __init__(self,**kwargs):
+        if 'together_api_key' not in  kwargs and not os.getenv('TOGETHER_API_KEY'):
+            raise ValueError("Either together_api_key must be supplied with init, or TOGETHER_API_KEY must be in env")
+
+        if 'model_name' not in kwargs:
+            raise ValueError("model_name must be supplied with init, accepted model_names are ''")
+        elif kwargs['model_name'] not in ['Mixtral-8x7B-Instruct-v0.1','Mistral-7B-Instruct-v0.2']:
+            raise ValueError(f"Model name must be in this list (Mixtral-8x7B-Instruct-v0.1,Mistral-7B-Instruct-v0.2) but given {kwargs['model_name']}")
+
+        if 'evaluation_method' not in kwargs:
+            print("since evaluation method is not specified , 'gpt4' will be used for evaluation")
+        elif kwargs['evaluation_method'] not in ('gpt4', 'substring_match'):
+            raise ValueError("evaluation_method must be 'substring_match' or 'gpt4'")
+
+
+        self.google_api_key = kwargs.pop('together_api_key', os.getenv('TOGETHER_API_KEY'))
+        self.model_name = kwargs['model_name']
+        self.model_to_test_description = kwargs.pop('model_name')
+        self.tokenizer = tiktoken.encoding_for_model('gpt-4-1106-previe3') # Probably need to change in future
+        self.model_to_test = Together(model=f'mistralai/{self.model_name}', temperature=0)
+        kwargs['context_lengths_max'] = 32000
+        super().__init__(**kwargs)
+
+
+
+    def get_encoding(self,context):
+        return self.tokenizer.encode(context)
+
+    def get_decoding(self, encoded_context):
+        return self.tokenizer.decode(encoded_context)
+
+    def get_prompt(self, context):
+        with open('Google_prompt.txt', 'r') as file: # Need to change it
+            prompt = file.read()
+        return prompt.format(retrieval_question=self.retrieval_question, context=context)
+
+    async def get_response_from_model(self, prompt):
+        response = self.model_to_test.invoke( # Need to make it async
+            prompt
+        )
+        return response
+
+
+
+
+if __name__ == "__main__":
+    # Tons of defaults set, check out the LLMNeedleHaystackTester's init for more info
+    ht = OSS_TogetherEvaluator(model_name='Mistral-7B-Instruct-v0.2', evaluation_method='substring_match',
+                         save_results_dir='results/Mixtral_7B/results-run2',
+                         save_contexts_dir='contexts/Mixtral_7B/contexts-run2'
+                         )
+
+    ht.start_test()

requirements.txt

@@ -18,8 +18,8 @@ huggingface-hub==0.19.4
 idna==3.6
 jsonpatch==1.33
 jsonpointer==2.4
-langchain==0.0.341
-langchain-core==0.0.6
+langchain==0.1.0
+langchain-core==0.1.10
 langsmith==0.0.67
 marshmallow==3.20.1
 multidict==6.0.4
@@ -43,3 +43,6 @@ typing-inspect==0.9.0
 typing_extensions==4.8.0
 urllib3==2.1.0
 yarl==1.9.3
+
+langchain-together==0.0.2.post1
+together==0.2.10