improve hyperparams

ai_single_response.py

@@ -115,8 +115,10 @@ def query_gpt_model(
     kparam=150,
     temp=0.75,
     top_p=0.65,
+    batch_size=64,
     verbose=False,
     use_gpu=False,
+    beams=2,
 ):
     """
     query_gpt_model - the main function that calls the model.
@@ -153,19 +155,22 @@ def query_gpt_model(
     # call the model
     print("\n... generating...")
     this_result = ai.generate(
-        n=1,
-        top_k=kparam,
-        batch_size=128,
-        # the prompt input counts for text length constraints
-        max_length=resp_length + pr_len,
-        min_length=resp_min + pr_len,
-        prompt=this_prompt,
-        temperature=temp,
-        top_p=top_p,
-        do_sample=True,
-        return_as_list=True,
-        use_cache=True,
-    )
+                                n=1,
+                                top_k=kparam,
+                                batch_size=batch_size,
+                                # the prompt input counts for text length constraints
+                                max_length=resp_length + pr_len,
+                                min_length=resp_min + pr_len,
+                                prompt=this_prompt,
+                                temperature=temp,
+                                top_p=top_p,
+                                top_k=kparam,
+                                do_sample=True,
+                                return_as_list=True,
+                                use_cache=True,
+                                num_beams=beams,
+                                no_repeat_ngram_size=2,
+                            )
     if verbose:
         print("\n... generated:\n")
         pp.pprint(this_result)  # for debugging

app.py

@@ -73,6 +73,7 @@ def ask_gpt(message: str):
         kparam=150,
         temp=0.75,
         top_p=0.65,
+        beams=4,
     )
     if basic_sc:
         cln_resp = symspeller(resp["out_text"], sym_checker=schnellspell)