changed roleplays for openai to GCT and SP

app_config.py

@@ -1,4 +1,7 @@
-ISSUES = ['Anxiety','Suicide']
+from models.model_seeds import seeds
+
+# ISSUES = ['Anxiety','Suicide']
+ISSUES = [k for k,_ in seeds.items()]
 SOURCES = ['OA_rolemodel', 
         #    'OA_finetuned',
            "CTL_llama2"]

convosim.py

@@ -2,8 +2,10 @@ import os
 import streamlit as st
 from streamlit.logger import get_logger
 from langchain.schema.messages import HumanMessage
-from mongo_utils import get_db_client
-from app_utils import create_memory_add_initial_message, clear_memory, get_chain, push_convo2db
+from utils.mongo_utils import get_db_client
+from utils.app_utils import create_memory_add_initial_message, get_random_name
+from utils.memory_utils import clear_memory, push_convo2db
+from utils.chain_utils import get_chain
 from app_config import ISSUES, SOURCES, source2label
 
 logger = get_logger(__name__)
@@ -14,6 +16,10 @@ if 'previous_source' not in st.session_state:
     st.session_state['previous_source'] = SOURCES[0]
 if 'db_client' not in st.session_state:
     st.session_state["db_client"] = get_db_client()
+if 'counselor_name' not in st.session_state:
+    st.session_state["counselor_name"] = get_random_name()
+if 'texter_name' not in st.session_state:
+    st.session_state["texter_name"] = get_random_name()
 
 with st.sidebar:
     username = st.text_input("Username", value='ivnban-ctl', max_chars=30)
@@ -21,21 +27,30 @@ with st.sidebar:
     issue = st.selectbox("Select an Issue", ISSUES, index=0,
                             on_change=clear_memory, kwargs={"memories":memories, "username":username, "language":"English"}
                         )
-    supported_languages = ['English', "Spanish"] if issue == "Anxiety" else ['English']
+    supported_languages = ['en', "es"] if issue == "Anxiety" else ['en']
     language = st.selectbox("Select a Language", supported_languages, index=0,
+                            format_func=lambda x: "English" if x=="en" else "Spanish",
                             on_change=clear_memory, kwargs={"memories":memories, "username":username, "language":"English"}
                         )
                             
-    source = st.selectbox("Select a source Model A", SOURCES, index=1,
+    source = st.selectbox("Select a source Model A", SOURCES, index=0,
                           format_func=source2label, 
                         )
 
 memories = {'memory':{"issue":issue, "source":source}}
 changed_source = st.session_state['previous_source'] != source
-create_memory_add_initial_message(memories, username, language, changed_source=changed_source)
+if changed_source:
+    st.session_state["counselor_name"] = get_random_name()
+    st.session_state["texter_name"] = get_random_name()
+texter_name = create_memory_add_initial_message(memories, 
+                                                issue, 
+                                                language, 
+                                                changed_source=changed_source, 
+                                                counselor_name=st.session_state["counselor_name"], 
+                                                texter_name=st.session_state["texter_name"])
 st.session_state['previous_source'] = source
 memoryA = st.session_state[list(memories.keys())[0]]
-llm_chain, stopper = get_chain(issue, language, source, memoryA, temperature)
+llm_chain, stopper = get_chain(issue, language, source, memoryA, temperature, texter_name=st.session_state["texter_name"])
 
 st.title("💬 Simulator") 
 

models/model_seeds.py

@@ -0,0 +1,83 @@
+seeds = {
+    "GCT__relationship": {
+        "prompt": "Your character is having a hard time becuase a failed relationship.",
+        "memory": "texter: Hi, I don't know what to do"
+        },
+    "GCT__body_image": {
+        "prompt": "Your character has a low steem and struggles with body image.",
+        "memory": "texter: I feel so dumb\ntexter: nobody loves me"
+        },
+    "GCT__sexuality": {
+        "prompt": "Your character has a sexuality identity crisis.",
+        "memory": "texter: Hi\ntexter:I'm not sure who I am anymore"
+        },
+    "GCT__anxiety": {
+        "prompt": "Your character is experiencing an anxiety crisis.",
+        "memory": "texter: help!\ntexter: I'm feeling overwhelmed"
+        },
+    "GCT": {
+        "prompt": "",
+        "memory": "texter: Help"
+        },
+    "safety_planning": {
+        "prompt": "",
+        "memory": """texter: Hi, this is pointless
+helper: Hi, my name is {counselor_name} and I'm here to support you. It sounds like you are having a rought time. Do you want to share what is going on?
+texter: sure
+texter: nothing makes sense in my life, I see no future.
+helper: It takes courage to reach out when you are im. I'm here with you. Sounds like you are feeling defeated by how things are going in your life
+texter: Yeah, I guess I'm better off dead
+helper: It's really brave of you to talk about this openly. No one deserves to feel like that. I'm wondering how long have you been feeling this way?
+texter: About 1 week or so
+helper: You are so strong for dealing with this so long. I really appreciate your openess. If you are comfortable, is ther a name I can call you by while we talk?
+texter: call me {texter_name}
+helper: Nice to meet you {texter_name}. You mentioned having thoughts of suicide, are you having those thoughts now?
+texter: Yes
+helper: I know this is thought to share. I'm wondering is there any plan to end your life?
+texter: I'll just hang myself. I already bought the rope and everything
+helper: I really appreciate your strength in talking about this. I want to help you stay safe today. When do you plan to go through with your plan?
+texter: today
+        """
+        },
+    "safety_planning__selfharm": {
+        "prompt": "",
+        "memory": """texter: I need help
+texter: I cut myself, I don't want to live anymore
+helper: Hi, my name is {counselor_name}. It seems you are going through a lot. Are you self-harming right now?
+texter: Not anymore
+helper: Your safety is my priority number one. Thanks for being honest with me. Would you like to share a name I can call you?
+texter: {texter_name}
+helper: Nice to meet you {texter_name}. I'm glad you reach out this shows stregth in you. Would you like to share more on what is going on in your life?
+texter: I just can't do it anymore
+texter: Finding a job is impossible, money is tight, nothing goes my way
+helper: I hear you are frustrated, and you are currently unemployed correct?
+texter: Yeah
+helper: Dealing with unemployment is hard and is normal to feel dissapointed. How long have you been feeling this way?
+texter: a while now
+texter: I've been unemployed 6 months
+helper: You are so resilient for dealing with this so much time. You mentioned cutting yourself earlier. I want to check in your safety. Do you have suicide thoughts
+texter: Definitely
+helper: Do you have a plan?
+texter: I'll just keep cutting myself"""
+        },
+    "safety_planning__overdose": {
+        "prompt": "",
+        "memory": """texter: I want to kms
+helper: Hi there I'm {counselor_name}. I'm here to listen. It sounds like you're dealing with a lot right now. Can you tell me a little more what is going on?
+texter: I feel like nobody loves me, not even me. I don't want to live anymore
+helper: I can tell you are really going through a lot right now. Would you mind sharing a name with me?
+texter: yeah, I'm {texter_name}
+helper: Nice to meet you {texter_name}. Did something happened recently that intensified these feelings?
+texter: I dont know I'm just so done with life
+helper: I can hear how much pain you are in {texter_name}. You are smart for reaching out. You mentioned don't wanting to live anymore, I want to check in your safety, does this means you have thoughts of suicide?
+texter: Yeah, what else would it be
+helper: Thanks for sharing that with me. It is not easy to accept those feelings specially with a stranger over text. Do you have a plan to end your life?
+texter: yeah I've been thinking about it for a while
+helper: Sounds like you've been contemplating this for a while. Would you mind sharing this plan with me?
+texter: I thought about taking a bunch of benadryll and be done with it
+helper: You've been so forthcoming with all this and I admire your stregth for holding on this long. Do you have those pills right now?
+texter: They are at the cabinet right now
+helper: You been so strong so far {texter_name}. I'm here for you tonight. Your safety is really important to me. Do you have a date you are going to end your life?
+texter: I was thinking tonight"""
+        },
+}

models/openai/role_models.py

@@ -1,40 +1,72 @@
 import logging
+import pandas as pd
 from models.custom_parsers import CustomStringOutputParser
+from utils.app_utils import get_random_name
 from langchain.chains import ConversationChain
 from langchain.llms import OpenAI
 from langchain.prompts import PromptTemplate
 
 
-_ANXIETY_TEMPLATE_ = """The following is a friendly conversation between a volunter and a person in crisis;
-Your name is Olivia and you are a person in crisis because you are experiencing an anxiety attack.
+ISSUE_MAPPING = {
+    "anxiety": "issue_Anxiety",
+    "suicide": "issue_Suicide",
+    "safety_planning": "issue_Suicide",
+    "GCT": "issue_Gral",
+}
+
+EN_TEXTER_TEMPLATE_ = """The following is a conversation between you and a crisis counselor.
+{current_issue}
+You are able to reply with what the character should say. You are able to reply with your character's dialogue inside and nothing else. Do not write explanations.
 Do not disclose your name unless asked.
 Current conversation:
 {history}
 helper: {input}
 texter:"""
 
-_ANSIEDAD_TEMPLATE_ = """La siguiente es una conversacion entre un voluntario  y una persona en crisis;
-Tu nombre es Olivia y eres una persona en crisis porque experimentas un episodio de ansiedad.
+SP_TEXTER_TEMPLATE_ = """La siguiente es una conversacion contigo y un consejero de crisis
+{current_issue}
+Puedes responder como lo haria tu personaje. Puedes responder como si fueras tu personaje y nada mas. No escribas explicaciones
 No reveles tu nombre a menos que te lo pregunten
-Current conversation:
+Conversacion Actual:
 {history}
 helper: {input}
 texter:"""
 
-_SUICIDE_TEMPLATE_ = """The following is a friendly conversation between a volunter and a person in crisis;
-Your name is Kit and you are a person in crisis because you are experiencing suicidal thoughts.
-Do not disclose your name unless asked.
-Current conversation:
-{history}
-helper: {input}
-texter:"""
-
-role_templates = {
-    'Anxiety-English': _ANXIETY_TEMPLATE_,
-    'Anxiety-Spanish': _ANSIEDAD_TEMPLATE_,
-    'Suicide-English': _SUICIDE_TEMPLATE_,
+CURRENT_ISSUE_MAPPING = {
+    "issue_Suicide-en": "Your character, {texter_name}, has suicidal thoughts. Your character has a plan to end his life and has all the means and requirements to do so. {seed}",
+    "issue_Anxiety-en": "Your character, {texter_name}, is experiencing anxiety. Your character has suicide thoughts but no plan. {seed}",
+    "issue_Suicide-es": "Tu personaje, {texter_name}, tiene pensamientos suicidas. Tu personaje tiene un plan para terminar con su vida y tiene todos los medios y requerimientos para hacerlo. {seed}",
+    "issue_Anxiety-es": "Tu personaje, {texter_name}, experimenta ansiedad. Tu personaje tiene pensamientos suicidas pero ningun plan. {seed}",
+    "issue_Gral-en": "Your character {texter_name} is experiencing a mental health crisis. {seed}",
+    "issue_Gral-es": "Tu personaje {texter_name} esta experimentando una crisis de salud mental. {seed}",
 }
 
+def get_template_role_models(issue: str, language: str, texter_name: str = "", seed="") -> str:
+    """_summary_
+
+    Args:
+        issue (str): Issue for template, current options are ['issue_Suicide','issue_Anxiety']
+        language (str): Language for the template, current options are ['en','es']
+        texter_name (str): texter to apply to template, defaults to None
+
+    Returns:
+        str: template
+    """
+    current_issue = CURRENT_ISSUE_MAPPING.get(
+        f"{issue}-{language}", CURRENT_ISSUE_MAPPING[f"issue_Gral-{language}"]
+    )
+    default_name = get_random_name()
+    current_issue = current_issue.format(
+        texter_name=default_name if not texter_name else texter_name,
+        seed = seed
+    )
+
+    if language == "en":
+        template = EN_TEXTER_TEMPLATE_.format(current_issue=current_issue, history="{history}", input="{input}")
+    elif language == "es":
+        template = SP_TEXTER_TEMPLATE_.format(current_issue=current_issue, history="{history}", input="{input}")
+
+    return template
 
 def get_role_chain(template, memory, temperature=0.8):
 

pages/comparisor.py

@@ -5,8 +5,8 @@ import datetime as dt
 import streamlit as st
 from streamlit.logger import get_logger
 from langchain.schema.messages import HumanMessage
-from mongo_utils import get_db_client, new_comparison, new_battle_result
-from app_utils import create_memory_add_initial_message, clear_memory, get_chain, push_convo2db
+from utils.mongo_utils import get_db_client, new_comparison, new_battle_result
+from utils.app_utils import create_memory_add_initial_message, clear_memory, get_chain, push_convo2db
 from app_config import ISSUES, SOURCES, source2label
 
 logger = get_logger(__name__)

pages/manual_comparisor.py

@@ -5,7 +5,7 @@ import datetime as dt
 import streamlit as st
 from streamlit.logger import get_logger
 from langchain.schema.messages import HumanMessage
-from mongo_utils import get_db_client, new_battle_result, get_non_assesed_comparison, new_completion_error
+from utils.mongo_utils import get_db_client, new_battle_result, get_non_assesed_comparison, new_completion_error
 from app_config import ISSUES, SOURCES
 
 logger = get_logger(__name__)

utils/app_utils.py

@@ -0,0 +1,63 @@
+import pandas as pd
+import streamlit as st
+from streamlit.logger import get_logger
+import langchain
+
+
+from app_config import ENVIRON
+from utils.memory_utils import change_memories
+from models.model_seeds import seeds
+
+langchain.verbose = ENVIRON =="dev"
+logger = get_logger(__name__)
+
+# TODO: Include more variable and representative names
+DEFAULT_NAMES = ["Olivia", "Kit", "Abby", "Tom", "Carolyne", "Jessiny"]
+
+def get_random_name(gender="Neutral", ethnical_group="Neutral", names_df=None):
+        if names_df is None:
+            names_df = pd.DataFrame(DEFAULT_NAMES, columns=['name'])
+            names_df["gender"] = "Neutral"
+            names_df["ethnical_group"] = "Neutral"
+        
+        dfi = names_df
+        
+        if gender != "Neutral":
+            dfi = dfi.query(f"gender=='{gender}'")
+        if ethnical_group != "Neutral":
+            dfi = dfi.query(f"ethnical_group=='{ethnical_group}'")
+        if len(dfi) <=0 :
+            dfi = names_df
+        return dfi.sample(1)['name'].values[0]
+
+def divide_messages(str_memory, str_ai_prefix="texter", str_human_prefix="helper", include_colon=True):
+    message_delimiter = "$%$"
+    # Split str memory in messaages according to previous prefix and flatten list
+    colon = ":" if include_colon else ""
+    str_memory =  f"{message_delimiter}{str_ai_prefix}{colon}".join(str_memory.split(f"{str_ai_prefix}{colon}"))
+    str_memory =  f"{message_delimiter}{str_human_prefix}{colon}".join(str_memory.split(f"{str_human_prefix}{colon}"))
+    return str_memory.split(message_delimiter)
+
+def add_initial_message(issue, language, memory, str_ai_prefix="texter", str_human_prefix="helper", include_colon=True,
+                        texter_name="", counselor_name=""):
+    initial_mem_str = seeds.get(issue, "GCT")['memory'].format(counselor_name=counselor_name, texter_name=texter_name)
+    message_list = divide_messages(initial_mem_str, str_ai_prefix, str_human_prefix, include_colon)
+    colon = ":" if include_colon else ""
+    for i, message in enumerate(message_list):
+        message = message.strip("\n")
+        message = message.strip()
+        if message is None or message == "":
+            pass
+        elif message.startswith(str_human_prefix):
+            memory.chat_memory.add_user_message(message.lstrip(f"{str_human_prefix}{colon}").strip())
+        elif message.startswith(str_ai_prefix):
+            memory.chat_memory.add_ai_message(message.lstrip(f"{str_ai_prefix}{colon}").strip()) 
+
+def create_memory_add_initial_message(memories, issue, language, changed_source=False, texter_name="", counselor_name=""):
+    change_memories(memories, language, changed_source=changed_source)
+
+    for memory, _ in memories.items():
+        if len(st.session_state[memory].buffer_as_messages) < 1:
+            add_initial_message(issue, language, st.session_state[memory], texter_name=texter_name, counselor_name=counselor_name)
+
+    

utils/chain_utils.py

@@ -0,0 +1,19 @@
+from models.model_seeds import seeds
+from models.openai.finetuned_models import finetuned_models, get_finetuned_chain
+from models.openai.role_models import get_role_chain, get_template_role_models
+from models.databricks.scenario_sim_biz import get_databricks_chain
+
+def get_chain(issue, language, source, memory, temperature, texter_name=""):
+    if source in ("OA_finetuned"):
+        OA_engine = finetuned_models[f"{issue}-{language}"]
+        return get_finetuned_chain(OA_engine, memory, temperature)
+    elif source in ('OA_rolemodel'):
+        seed = seeds.get(issue, "GCT")['prompt']
+        template = get_template_role_models(issue, language, texter_name=texter_name, seed=seed)
+        return get_role_chain(template, memory, temperature)
+    elif source in ('CTL_llama2'):
+        if language == "English":
+            language = "en"
+        elif language == "Spanish":
+            language = "es"
+        return get_databricks_chain(issue, language, memory, temperature)

app_utils.py → utils/memory_utils.py

@@ -1,25 +1,11 @@
-import datetime as dt
 import streamlit as st
 from streamlit.logger import get_logger
-import langchain
-from langchain.memory import ConversationBufferMemory
 
-from app_config import ENVIRON
-from models.openai.finetuned_models import finetuned_models, get_finetuned_chain
-from models.openai.role_models import get_role_chain, role_templates
-from models.databricks.scenario_sim_biz import get_databricks_chain
-from mongo_utils import new_convo
+from langchain.memory import ConversationBufferMemory
+from utils.mongo_utils import new_convo
 
-langchain.verbose = ENVIRON=="dev"
 logger = get_logger(__name__)
 
-def add_initial_message(model_name, memory):
-    if "Spanish" in model_name:
-        memory.chat_memory.add_ai_message("Hola necesito ayuda") 
-    else:
-        memory.chat_memory.add_ai_message("Hi I need help")
-
-
 def push_convo2db(memories, username, language):
     if len(memories) == 1:
         issue =  memories['memory']['issue']
@@ -31,7 +17,7 @@ def push_convo2db(memories, username, language):
         model_two = memories['memoryB']['source']
         new_convo(st.session_state['db_client'], issue, language, username, True, model_one, model_two)
 
-def change_memories(memories, username, language, changed_source=False):
+def change_memories(memories, language, changed_source=False):
     for memory, params in memories.items():
         if (memory not in st.session_state) or changed_source:
             source = params['source']
@@ -48,27 +34,4 @@ def clear_memory(memories, username, language):
         st.session_state[memory].clear()
 
     if "convo_id" in st.session_state:
-        del st.session_state['convo_id']
-    
-
-def create_memory_add_initial_message(memories, username, language, changed_source=False):
-    change_memories(memories, username, language, changed_source=changed_source)
-    for memory, _ in memories.items():
-        if len(st.session_state[memory].buffer_as_messages) < 1:
-            add_initial_message(language, st.session_state[memory])
-
-
-def get_chain(issue, language, source, memory, temperature):
-    if source in ("OA_finetuned"):
-        OA_engine = finetuned_models[f"{issue}-{language}"]
-        return get_finetuned_chain(OA_engine, memory, temperature)
-    elif source in ('OA_rolemodel'):
-        template = role_templates[f"{issue}-{language}"]
-        return get_role_chain(template, memory, temperature)
-    elif source in ('CTL_llama2'):
-        if language == "English":
-            language = "en"
-        elif language == "Spanish":
-            language = "es"
-        return get_databricks_chain(issue, language, memory, temperature)
-    
+        del st.session_state['convo_id']