Merge pull request #18 from sotopia-lab/feature/add_finetuned_model

Add new model version for gpt-3.5-turbo-finetuned

.gitignore

@@ -141,5 +141,3 @@ deprecated/*
 
 #backup
 backup/*
-
-scripts/*

examples/generate_scenarios.py

@@ -1,13 +1,15 @@
 import ast
 import asyncio
+import json
+import random
 from typing import Any, cast
 
 import pandas as pd
 import typer
 from experiment_eval import _sample_env_agent_combo_and_push_to_db
 from redis_om import Migrator
 
-from sotopia.database import EnvironmentProfile
+from sotopia.database import EnvAgentComboStorage, EnvironmentProfile
 from sotopia.database.persistent_profile import RelationshipType
 from sotopia.generation_utils import (
     LLM_Name,
@@ -37,11 +39,15 @@ def add_env_profiles(
 def check_existing_envs(
     env_profile: dict[str, Any], existing_envs: pd.DataFrame
 ) -> bool:
-    if (
-        env_profile["scenario"] in existing_envs["scenario"].to_list()
-        and str(env_profile["agent_goals"])
-        in existing_envs["agent_goals"].to_list()
-    ):
+    try:
+        if (
+            env_profile["scenario"] in existing_envs["scenario"].to_list()
+            and str(env_profile["agent_goals"])
+            in existing_envs["agent_goals"].to_list()
+        ):
+            return False
+    except KeyError:
+        print(env_profile)
         return False
     return True
 
@@ -50,7 +56,7 @@ def generate_newenv_profile(
     num: int,
     gen_model: LLM_Name = "gpt-4-turbo",
     temperature: float = 0.5,
-    type: str = "mutual_friend",
+    type: str = "craigslist_bargains",
 ) -> pd.DataFrame:
     env_profile_list = []  # type: ignore
     existing_envs = pd.read_csv(
@@ -70,6 +76,22 @@ def generate_newenv_profile(
             }
             if check_existing_envs(env_profile, existing_envs):
                 env_profile_list.append(env_profile)
+    elif type == "craigslist_bargains":
+        while len(env_profile_list) < num:
+            scenario, social_goals = asyncio.run(
+                generate_craigslist_bargains_envs()
+            )
+            env_profile = {
+                "codename": f"craigslist_bargains_{len(env_profile_list)+10}",
+                "scenario": scenario,
+                "agent_goals": social_goals,
+                "relationship": RelationshipType.stranger,
+                "age_constraint": "[(18, 80), (18, 80)]",
+                "occupation_constraint": None,
+                "source": "craigslist_bargains",
+            }
+            if check_existing_envs(env_profile, existing_envs):
+                env_profile_list.append(env_profile)
     else:
         raise NotImplementedError("Only mutual_friend is supported for now")
     return pd.DataFrame(env_profile_list)
@@ -116,5 +138,59 @@ def auto_generate_scenarios(
     Migrator().run()
 
 
+@app.command()
+def clean_env_wo_combos() -> None:
+    """
+    Function to clean up env-agent combos in the database
+    """
+    env_agent_combos = list(EnvAgentComboStorage.all_pks())
+    envs_id_in_combos = set(
+        [
+            EnvAgentComboStorage.get(env_agent_combo).env_id
+            for env_agent_combo in env_agent_combos
+        ]
+    )
+    envs = list(EnvironmentProfile.all_pks())
+    for env in envs:
+        if env not in envs_id_in_combos:
+            EnvironmentProfile.delete(env)
+
+
+@app.command()
+def upload_env_profiles(
+    filepath: str = "./data/all_environment_profile.json",
+) -> None:
+    """
+    Function to upload environment profiles from json file
+    The json file format is a direct dump from the database
+    """
+    env_profile_list = []
+    existing_envs = pd.read_csv(
+        "./data/env_profiles_v1.csv"
+    )  # TODO: find a better way to deal with this
+    current_envs = json.load(open(filepath, "r"))
+    for key in current_envs:
+        env_profile = current_envs[key]
+        if env_profile and check_existing_envs(env_profile, existing_envs):
+            del env_profile["pk"]
+            env_profile_list.append(env_profile)
+    env_profiles = add_env_profiles(env_profile_list)
+    print("New env profiles added to database:")
+    print(len(env_profiles))
+
+    count = 0
+    for env_profile in env_profiles:
+        assert env_profile.pk is not None
+        try:
+            _sample_env_agent_combo_and_push_to_db(env_profile.pk)
+            count += 1
+        except:
+            EnvironmentProfile.delete(env_profile.pk)
+            pass
+    print(f"New env-agent combo added to database: {count}")
+
+    Migrator().run()
+
+
 if __name__ == "__main__":
     app()

scripts/evaluate_finetuned_MF.sh

@@ -0,0 +1,16 @@
+MODEL_NAME_1=gpt-3.5-turbo-ft-MF
+MODEL_NAME_2=gpt-3.5-turbo
+
+python examples/experiment_eval.py \
+ --gin_file sotopia_conf/generation_utils_conf/generate.gin \
+ --gin_file sotopia_conf/server_conf/server.gin \
+ --gin_file sotopia_conf/run_async_server_in_batch.gin \
+ "--gin.ENV_IDS=['01H7VFHPKA2GGPPNVJWV967HZC', '01H7VFHPHWA2CYG7BC82NS4XH1', '01H7VFHPH567HKQRE0C745KH9C', '01H7VFHPMS6AJY0PFGGCFFK5GX', '01H7VFHPJKR16MD1KC71V4ZRCF', '01H7VFHPQ1712DHGTMPQFTXH02', '01H7VFHPP9SPQ8W6583JFZ7HZC', '01H7VFHPM3NVVKSGCCB4S10465', '01H7VFHPGABSWQXTACCC8C3X2F', '01H7VFHPNHZ2YYRHP0GXARD550']" \
+ "--gin.AGENT1_MODEL=\"${MODEL_NAME_1}\"" \
+ "--gin.AGENT2_MODEL=\"${MODEL_NAME_2}\"" \
+ '--gin.BATCH_SIZE=1' \
+ '--gin.TAG="finetuned_gpt3.5_gpt3.5ft_MF"' \
+ '--gin.TAG_TO_CHECK_EXISTING_EPISODES="finetuned_gpt3.5_gpt3.5ft_MF"' \
+ '--gin.PUSH_TO_DB=True' \
+ '--gin.VERBOSE=False' \
+ '--gin.LITE=False' \

scripts/evaluate_finetuned_full.sh

@@ -0,0 +1,15 @@
+MODEL_NAME=gpt-3.5-turbo-finetuned
+
+python examples/experiment_eval.py \
+ --gin_file sotopia_conf/generation_utils_conf/generate.gin \
+ --gin_file sotopia_conf/server_conf/server.gin \
+ --gin_file sotopia_conf/run_async_server_in_batch.gin \
+ '--gin.ENV_IDS=[]' \
+ "--gin.AGENT1_MODEL=\"${MODEL_NAME}\"" \
+ "--gin.AGENT2_MODEL=\"${MODEL_NAME}\"" \
+ '--gin.BATCH_SIZE=5' \
+ '--gin.TAG="finetuned_gpt3.5"' \
+ '--gin.TAG_TO_CHECK_EXISTING_EPISODES="finetuned_gpt3.5"' \
+ '--gin.PUSH_TO_DB=True' \
+ '--gin.VERBOSE=False' \
+ '--gin.LITE=False' \

exp_scripts/exp_instruction.md → scripts/exp_instruction.md

@@ -1,5 +1,12 @@
+# Agent vs Storyteller Scripts
+
+### Basic Scripts
 Here are some of the script for running {gpt-3.5-turbo, mixtral-7b-moe} under {normal interaction, omniscient interaction, script generation} mode in {normal, lite} setting.
 If you need to run all interaction mode, you can use `run_all.sh`, the usage is `Usage: ./run_all.sh <model_name> <tag_base> <lite>`. For example, `./run_all.sh gpt-3.5-turbo exp0128 True`. You may find model_name in `LLM_Name`, and currently we are using `mistralai/Mixtral-8x7B-Instruct-v0.1` and `gpt-3.5-turbo`.
 If you want to run mode separately, you can use `run_interaction.sh` or `run_script_full.sh`.
 After running the above script, you may specify tags and fix those error episodes using `./fix_missing_episodes_with_tag.sh`.
 Current `fix_missing_episodes_with_tag.py` first detects erroneous episodes, delete them and regenerate them.
+
+### Fine-tuning
+
+* `evaluate_finetuned_full.sh`: evaluate the fine-tuned model (gpt-3.5 finetuned on the full dataset) on the sotopia lite setting.

sotopia/generation_utils/generate.py

@@ -50,6 +50,8 @@
     "togethercomputer/llama-2-70b-chat",
     "togethercomputer/mpt-30b-chat",
     "gpt-3.5-turbo",
+    "gpt-3.5-turbo-finetuned",
+    "gpt-3.5-turbo-ft-MF",
     "text-davinci-003",
     "gpt-4",
     "gpt-4-turbo",
@@ -290,11 +292,11 @@ def _type(self) -> str:
         return "str"
 
 
-def _return_fixed_model_version(
-    model_name: Literal["gpt-3.5-turbo", "gpt-4", "gpt-4-turbo"]
-) -> str:
+def _return_fixed_model_version(model_name: LLM_Name) -> str:
     return {
         "gpt-3.5-turbo": "gpt-3.5-turbo-0613",
+        "gpt-3.5-turbo-finetuned": "ft:gpt-3.5-turbo-0613:academicscmu::8nY2zgdt",
+        "gpt-3.5-turbo-ft-MF": "ft:gpt-3.5-turbo-0613:academicscmu::8nuER4bO",
         "gpt-4": "gpt-4-0613",
         "gpt-4-turbo": "gpt-4-1106-preview",
     }[model_name]
@@ -313,7 +315,7 @@ def obtain_chain(
     Using langchain to sample profiles for participants
     """
     match model_name:
-        case "gpt-3.5-turbo" | "gpt-4" | "gpt-4-turbo":
+        case "gpt-3.5-turbo" | "gpt-4" | "gpt-4-turbo" | "gpt-3.5-turbo-finetuned" | "gpt-3.5-turbo-ft-MF":
             human_message_prompt = HumanMessagePromptTemplate(
                 prompt=PromptTemplate(
                     template=template,
@@ -781,7 +783,6 @@ async def agenerate_action(
                 Your action should follow the given format:
                 {format_instructions}
             """
-
         return await agenerate(
             model_name=model_name,
             template=template,