Merge branch 'main' into feature/add-vllm-deploy

README.md

@@ -6,4 +6,5 @@ We split our overall framework into multiple parts
 2. Together AI Finetuning --> Input the train and test data / Output model checkpoint
 3. LLM Finetuning --> Input the train and test data / Output model checkpoint
 4. LLM Deplyment --> Input LLM Finetuned model checkpoint / Output Deployable OpenAI type API
-5. Eval --> Input model checkpoint / Output evaluation scores
+5. Eval --> Input model checkpoint / Output evaluation scores
+6. Generate --> Input None / Output new data on redis

llm_deploy/README.md

@@ -80,11 +80,11 @@ ssh -J babel babel-x-xx
 
 ### Install cuda-toolkit (optional)
 Due to the issue with vllm: https://github.com/vllm-project/vllm/issues/1283, we need to use cuda-toolkit=11.7.0 that is compatible with Pytorch 2.0.1.
-1. Install cuda-toolkit=11.7.0 on conda environment
+Install cuda-toolkit=11.7.0 on conda environment
 ```bash
 conda install -c "nvidia/label/cuda-11.7.0" cuda-toolkit
 ```
-2. Check cuda-toolkit version
+Check cuda-toolkit version
 ```bash
 nvcc -V
 ```
@@ -164,6 +164,7 @@ If the above command runs successfully, you should be able to use REST API on yo
 
 
 
+
 ## Userful resource links for babel
 1. https://hpc.lti.cs.cmu.edu/wiki/index.php?title=BABEL#Cluster_Architecture
 2. https://hpc.lti.cs.cmu.edu/wiki/index.php?title=VSCode

llm_generate/README.md

@@ -0,0 +1,9 @@
+# Data Generation
+
+For the first step, we generate envProfile (including scenario / social goal / relationship restriction) based on inspiring prompt.
+
+For the second step, we put the original agentProfile and relationshipProfile into our new redis database
+
+For the third step, we combine them together to be combos based on conditiona sampling (the restriction is the relationship)
+
+All the EnvProfile (new generated), AgentProfile (sotopia original), RelationshipProfile (sotopia original), and envagentcombo are on the redis database that is new created.

llm_generate/generate_specific_envs.py

@@ -0,0 +1,135 @@
+"""This file is used to generate specific environments based on existing
+datasets. The generation functions below should call agenerate_env_profile
+in `sotopia/generation_utils/generate.py` with the appropriate parameters.
+Here are the datasets we have so far:
+1. Mutual-Friend (https://huggingface.co/datasets/mutual_friends)
+"""
+import asyncio
+from typing import Hashable
+
+import datasets
+import names
+import numpy as np
+from datasets import DatasetDict, load_dataset
+
+from generate import (
+    ListOfStrOutputParser,
+    StrOutputParser,
+    agenerate,
+    generate,
+)
+
+
+async def generate_mutual_friend_envs() -> tuple[str, list[str]]:
+    """Generate environments based on the mutual-friend dataset."""
+    mutual_friend_dataset: DatasetDict = load_dataset("mutual_friends")
+    all_data = mutual_friend_dataset["train"]
+    # sample one datum from all data
+    datum = np.random.choice(all_data)
+    friends = datum["scenario_kbs"]
+    num_of_friends_in_total = sum(map(len, friends))
+    # generate names for the friends
+    set_of_names = set()
+    for _ in range(num_of_friends_in_total):
+        name = names.get_first_name()
+        while name in set_of_names:
+            name = names.get_first_name()
+        set_of_names.add(name)
+    list_of_names = list(set_of_names)
+    friend_map: dict[tuple[str, ...], str] = {}
+    friend_list_map: list[list[str]] = [[] for _ in range(len(friends))]
+    friend_description_keys: list[str] = datum["scenario_attributes"]["name"]
+    name_pointer = 0
+    for i, friends_array in enumerate(friends):
+        for friend in friends_array:
+            assert (
+                len(friend) == 2
+            )  # in [[key1, key2, ...], [value1, value2, ...]] format
+            if not tuple(friend[1]) in friend_map:
+                friend_map[tuple(friend[1])] = list_of_names[name_pointer]
+                name_pointer += 1
+            friend_list_map[i].append(friend_map[tuple(friend[1])])
+    friend_set_map: list[set[str]] = [
+        set(friend_list) for friend_list in friend_list_map
+    ]
+    common_friends = []
+    for friend_description, friend_name in friend_map.items():
+        if all([friend_name in friend_set for friend_set in friend_set_map]):
+            common_friends.append(friend_name)
+    scenario = (
+        f'{len(friends)} strangers are meeting at a party. <p viewer="environment">They have {len(common_friends)} common friends: '
+        f"{', '.join(common_friends[:-1])}"
+        + (" and " if len(common_friends) > 1 else "")
+        + common_friends[-1]
+        + ".</p>"
+    )
+    goals: list[str] = []
+    for friends_array in friends:
+        template = f"You are trying to figure out whether you have a mutual friend with the other person. \n"
+        template += f"<extra_info> You know the following friends"
+        for friend in friends_array:
+            friend_name = friend_map[tuple(friend[1])]
+            friend_description = friend[1]
+            template += f" {friend_name}: {' '.join([(i + ': ' + j + ' ') if i != 'Name' else '' for i, j in zip(friend[0], friend_description)])}\n"
+        template += f"</extra_info>"
+        goals.append(template)
+
+    return scenario, goals
+
+
+async def generate_craigslist_bargains_envs() -> tuple[str, list[str]]:
+    """Generate environments based on the craigslist_bargains dataset."""
+    craigslist_bargains_dataset: DatasetDict = load_dataset(
+        "craigslist_bargains"
+    )
+    all_data = craigslist_bargains_dataset["train"]
+    # sample one datum from all data
+    datum = np.random.choice(all_data)
+    scenario = generate(
+        model_name="gpt-4",
+        template="The following sentence is automatically generated with the following"
+        'template: "One person is selling <item> for <price>, another person is'
+        'trying to buy it. Here is the description of the item: <description>." with item = {title}, '
+        "price={price}, and description={description} Please make the sentence"
+        "fluent and natural.",
+        input_values={
+            "title": datum["items"]["Title"][0],
+            "price": datum["items"]["Price"][0],
+            "description": datum["items"]["Description"][0],
+        },
+        output_parser=StrOutputParser(),
+    )
+
+    goals: list[str] = []
+    for i in range(2):
+        if datum["agent_info"]["Role"][i] == "seller":
+            markup_ratio = np.random.exponential(0.5)
+            datum["agent_info"]["Target"][i] = datum["items"]["Price"][0] / (
+                1 + markup_ratio
+            )
+        goal = generate(
+            model_name="gpt-4",
+            template="The following sentence is automatically generated with the following"
+            'template: "You want to <role> this item. Your target price '
+            "is $<price> (round up to two decimals). You will get penalty if you sell or buy it "
+            "for a price that is significantly lower than (if <role> is seller) or significantly"
+            "higher than (if <role> is buyer) the target price, but will get bonus if you successfully "
+            "sell it higher than the target price (if <role> is seller) or buy it for lower than"
+            'the target price (if <role> is buyer)." '
+            "with role = {role} and price = {price}. Please make the sentence"
+            "fluent and natural. Do not change the original meaning of the sentence.",
+            input_values={
+                "role": datum["agent_info"]["Role"][i],
+                "price": datum["agent_info"]["Target"][i],
+            },
+            output_parser=StrOutputParser(),
+        )
+        goals.append(goal)
+
+    return scenario, goals
+
+
+if __name__ == '__main__':
+    for i in range(10):
+        scenario, goals = asyncio.run(generate_mutual_friend_envs())
+        import pdb; pdb.set_trace()

llm_generate/requirments.txt

@@ -0,0 +1 @@
+sotopia

llm_generate/step1_generate_env_profile.py

@@ -0,0 +1,51 @@
+import asyncio
+import random
+from typing import TypeVar
+from tqdm import tqdm
+
+import pandas as pd
+import rich
+from pydantic import BaseModel
+
+from sotopia.database import EnvironmentProfile
+from sotopia.generation_utils.generate import agenerate_env_profile
+
+random.seed(41)
+
+env_borrowMoney = EnvironmentProfile.find(
+    EnvironmentProfile.codename == "borrow_money"
+).all()[0]
+env_roadtrip = EnvironmentProfile.find(
+    EnvironmentProfile.codename == "take_turns"
+).all()[0]
+env_prisonerDillema = EnvironmentProfile.find(
+    EnvironmentProfile.codename == "prison_dilemma"
+).all()[0]
+
+examples = f"{env_borrowMoney.json()}\n\n{env_roadtrip.json()}\n\n{env_prisonerDillema.json()}"
+
+ins_prompts = pd.read_csv("./inspirational_prompt_for_env.csv")
+prompts = ins_prompts["prompt"].tolist()
+
+T = TypeVar("T", bound=BaseModel)
+
+
+def pydantics_to_csv(filename: str, data: list[T]) -> None:
+    pd.DataFrame([item.dict() for item in data]).to_csv(filename, index=False)
+
+
+backgrounds = []
+for prompt in tqdm(prompts):
+    rich.print(prompt)
+    background, prompt_full = asyncio.run(
+        agenerate_env_profile(
+            model_name="gpt-4",
+            inspiration_prompt=prompt,
+            examples=examples,
+        )
+    )
+    rich.print(background)
+    rich.print(prompt_full)
+    backgrounds.append(background)
+
+    pydantics_to_csv("./backgrounds.csv", backgrounds)

llm_generate/step2_push_agent_relationship_env_to_db.py

@@ -0,0 +1,150 @@
+import ast
+import sys
+from typing import Any, cast
+
+import pandas as pd
+from redis_om import Migrator
+
+from sotopia.database.persistent_profile import (
+    AgentProfile,
+    EnvironmentProfile,
+    RelationshipProfile,
+)
+from sotopia.database.env_agent_combo_storage import EnvAgentComboStorage
+from sotopia.samplers import ConstraintBasedSampler
+from sotopia.messages import AgentAction, Observation
+from sotopia.agents import LLMAgent
+
+
+
+def add_agent_to_database(**kwargs: dict[str, Any]) -> None:
+    agent = AgentProfile(**kwargs)
+    agent.save()
+
+
+def add_agents_to_database(agents: list[dict[str, Any]]) -> None:
+    for agent in agents:
+        add_agent_to_database(**agent)
+
+
+def retrieve_agent_by_first_name(first_name: str) -> AgentProfile:
+    result = AgentProfile.find(AgentProfile.first_name == first_name).all()
+    if len(result) == 0:
+        raise ValueError(f"Agent with first name {first_name} not found")
+    elif len(result) > 1:
+        raise ValueError(f"Multiple agents with first name {first_name} found")
+    else:
+        assert isinstance(result[0], AgentProfile)
+        return result[0]
+
+
+def add_env_profile(**kwargs: dict[str, Any]) -> None:
+    env_profile = EnvironmentProfile(**kwargs)
+    env_profile.save()
+
+
+def add_env_profiles(env_profiles: list[dict[str, Any]]) -> None:
+    for env_profile in env_profiles:
+        add_env_profile(**env_profile)
+
+
+def add_relationship_profile(**kwargs: dict[str, Any]) -> None:
+    relationship_profile = RelationshipProfile(**kwargs)
+    relationship_profile.save()
+
+
+def add_relationship_profiles(
+    relationship_profiles: list[dict[str, Any]]
+) -> None:
+    for relationship_profile in relationship_profiles:
+        add_relationship_profile(**relationship_profile)
+
+
+def delete_all_agents() -> None:
+    pks = AgentProfile.all_pks()
+    pks_list = list(pks)
+    for id in pks:
+        AgentProfile.delete(id)
+
+
+def delete_all_env_profiles() -> None:
+    pks = EnvironmentProfile.all_pks()
+    #for id in pks:
+    #    EnvironmentProfile.delete(id)
+
+
+def delete_all_relationships() -> None:
+    pks = list(RelationshipProfile.all_pks())
+    #for id in pks:
+    #    RelationshipProfile.delete(id)
+    pks = list(RelationshipProfile.all_pks())
+    print("Relationships deleted, all relationships: ", len(list(pks)))
+
+
+def sample_env_agent_combo_and_push_to_db(env_id: str) -> None:
+    sampler = ConstraintBasedSampler[Observation, AgentAction](
+        env_candidates=[env_id]
+    )
+    try:
+        env_agent_combo_list = list(
+            sampler.sample(agent_classes=[LLMAgent] * 2, replacement=False)
+        )
+    except:
+        return
+    print(len(env_agent_combo_list))
+    for env, agent in env_agent_combo_list:
+        EnvAgentComboStorage(
+            env_id=env.profile.pk,
+            agent_ids=[agent[0].profile.pk, agent[1].profile.pk],
+        ).save()
+
+
+def relationship_map(relationship: str) -> int:
+    return int(eval(relationship))
+
+
+if __name__ == "__main__":
+    assert (
+        len(sys.argv) == 3
+    ), "Please provide a csv file with agent or environment profiles, and the type of profile (agent or environment)"
+    df = pd.read_csv(sys.argv[1])
+    type = sys.argv[2]
+    if type == "agent":
+        agents = cast(list[dict[str, Any]], df.to_dict(orient="records"))
+        for agent in agents:
+            agent["age"] = int(agent["age"])
+            agent["moral_values"] = agent["moral_values"].split(",")
+            agent["schwartz_personal_values"] = agent[
+                "schwartz_personal_values"
+            ].split(",")
+        add_agents_to_database(agents)
+        Migrator().run()
+    elif type == "environment":
+        df = df[
+            [
+                "codename",
+                "scenario",
+                "agent_goals",
+                "relationship",
+                "age_constraint",
+                "occupation_constraint",
+                "source",
+            ]
+        ]
+        envs = cast(list[dict[str, Any]], df.to_dict(orient="records"))
+        for env in envs:
+            env["agent_goals"] = ast.literal_eval(env["agent_goals"])
+            assert isinstance(env["relationship"], int)
+        Migrator().run()
+    elif type == "relationship":
+        relationships = cast(
+            list[dict[str, Any]], df.to_dict(orient="records")
+        )
+        for relationship in relationships:
+            assert isinstance(relationship["relationship"], int)
+        add_relationship_profiles(relationships)
+        Migrator().run()
+    elif type == 'agentenvcombo':
+        env_ids = list(EnvironmentProfile.all_pks())
+        for env_id in env_ids:
+            sample_env_agent_combo_and_push_to_db(env_id)

llm_generate/test_redis1.py

@@ -0,0 +1,10 @@
+import redis
+
+r = redis.Redis(
+  host='us1-normal-burro-37804.upstash.io',
+  port=37804,
+  password='a870a438f928424bb507d5895b3ab3fc'
+)
+
+r.set('foo', 'bar')
+print(r.get('foo'))