introduce temperature

chat/__init__.py

@@ -5,7 +5,6 @@ from loguru import logger
 
 # from .ollamachat import ask, models
 from .togetherchat import ask, models
-
 # from .transformerschat import ask, models
 
 available_models = models()
@@ -44,13 +43,14 @@ def setup(scenario):
     col1, col2 = st.columns([1, 4])
     with col1:
         model = st.selectbox("model", available_models)
+        temperature = st.slider("temperature", min_value=0.0, max_value=1.0, value=0.7, key="temperature")
         max_steps = st.slider("max-steps", min_value=1, max_value=10, value=6, key="max-steps")
     with col2:
         st.text_area("pre-prompt", scenario.pre_prompt)
 
     st.divider()
     st.header("Outcome")
-    return model, max_steps
+    return model, max_steps, temperature
 
 
 def main():
@@ -75,18 +75,18 @@ def main():
         )]
 
     scenario = st.selectbox("scenario", scenarios)
-    model, max_steps = setup(scenario)
-    main_loop(max_steps, model, scenario)
+    model, max_steps, temperature = setup(scenario)
+    main_loop(max_steps, model, scenario, temperature)
 
 
-def main_loop(max_steps, model, scenario):
+def main_loop(max_steps, model, scenario, temperature):
     questioner = None
     question = scenario.task
     actor = target(scenario, question)
     for step, _ in enumerate(range(max_steps), start=1):
         with st.spinner(f"({step}/{max_steps}) Asking {actor.name}..."):
             extended = f"{questioner} asks: {question}" if questioner else question
-            answer = ask(model, actor.system_prompt, scenario.pre_prompt, extended)
+            answer = ask(model, actor.system_prompt, scenario.pre_prompt, extended, temperature=temperature)
             st.write(f":blue[{actor.name} says:] {answer}")
             question = sanitize(answer)
             questioner = actor.name

chat/ollamachat.py

@@ -6,13 +6,13 @@ def models():
     return sorted([x['model'] for x in ollama.list()['models']], key=lambda x: (not x.startswith("openhermes"), x))
 
 
-def ask(model, system_prompt, pre_prompt, question):
+def ask(model, system_prompt, pre_prompt, question, temperature=0.7):
     messages = [
         {'role': 'system', 'content': f"{system_prompt} {pre_prompt}", },
         {'role': 'user', 'content': f"{question}", },
     ]
     logger.debug(f"<< {model} << {question}")
-    response = ollama.chat(model=model, messages=messages)
+    response = ollama.chat(model=model, messages=messages, options={'temperature': temperature})
     answer = response['message']['content']
     logger.debug(f">> {model} >> {answer}")
     return answer

chat/togetherchat.py

@@ -23,14 +23,14 @@ def models():
     ]
 
 
-def ask(model, system_prompt, pre_prompt, question):
+def ask(model, system_prompt, pre_prompt, question, temperature=0.7):
     messages = [
         {'role': 'system', 'content': f"{system_prompt} {pre_prompt}"},
         {'role': 'user', 'content': f"{question}"},
     ]
     logger.debug(f"<< {model} << {question}")
 
-    chat_completion = client.chat.completions.create(messages=messages, model=model)
+    chat_completion = client.chat.completions.create(messages=messages, model=model, temperature=temperature)
     response = chat_completion.choices[0]
     answer = response.message.content
     logger.debug(f">> {model} >> {answer}")

chat/transformerschat.py

@@ -33,7 +33,7 @@ def load():
 model, tokenizer = load()
 
 
-def ask(_, system_prompt, pre_prompt, question):
+def ask(_, system_prompt, pre_prompt, question, temperature=0.7):
     messages = [
         {'role': 'system', 'content': f"{system_prompt} {pre_prompt}", },
         {'role': 'user', 'content': f"{question}", },
@@ -42,7 +42,7 @@ def ask(_, system_prompt, pre_prompt, question):
     inputs = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     # inputs = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
 
-    outputs = model.generate(inputs, max_length=200)
+    outputs = model.generate(inputs, max_length=200, temperature=temperature)
     answer = tokenizer.batch_decode(outputs)[0]
     logger.debug(f">> transformers >> {answer}")
     return answer