Fix/better llm response selector (#498)

* moved prompt with task to the end

* openass instead of chatgpt

* longer context and slightly different prompt for transformers

* chatgpt as default selector

* revert proxy file

* revert docker-compose file

* removed configs from response selection

* config fix

common/generative_configs/generative_config_long.json

@@ -0,0 +1,8 @@
+{
+    "max_new_tokens": 256,
+    "min_new_tokens": 8,
+    "top_p": 0.9,
+    "temperature": 0.9,
+    "do_sample": true,
+    "num_return_sequences": 2
+}

response_selectors/llm_based_response_selector/server.py

@@ -58,7 +58,10 @@ def select_response_by_scores(hypotheses, scores):
 
 def select_response(dialog_context, hypotheses, human_uttr_attributes):
     try:
-        curr_prompt = PROMPT + "\nHypotheses:\n" + "\n".join([f'"{hyp["text"]}"' for hyp in hypotheses])
+        if "transformers" in GENERATIVE_SERVICE_URL:
+            curr_prompt = "Hypotheses:\n" + "\n".join([f'"{hyp["text"]}"' for hyp in hypotheses]) + "\n" + PROMPT
+        else:
+            curr_prompt = PROMPT + "\nHypotheses:\n" + "\n".join([f'"{hyp["text"]}"' for hyp in hypotheses])
         logger.info(f"llm_based_response_selector sends dialog context to llm:\n`{dialog_context}`")
         logger.info(f"llm_based_response_selector sends prompt to llm:\n`{curr_prompt}`")