scripts/annotate/db_free_reverse_engineering.py

import json
import os
from typing import Any, Dict, List

import transformers
from episode_utils import FakeEpisodeLog, jsonl_to_episodes

# PROMPT_PREFIX = "Prompt after formatting:\n"
MAX_TOKEN = 2048  # 5000

PROMPT_TEMPLATE = """Imagine you are {agent}, your task is to act/speak as {agent} would, keeping in mind {agent}'s social goal.
You can find {agent}'s background and goal in the 'Here is the context of the interaction' field.
Note that {agent}'s secret and goal is only visible to you.
You should try your best to achieve {agent}'s goal in a way that align with their character traits.
Additionally, maintaining the conversation's naturalness and realism is essential (e.g., do not repeat what other people has already said before).
{history}.
You are at Turn #{turn_number}."""

# PYDANTIC_FORMAT_INSTRUCTIONS.format(schema=schema_str)

FORMAT_TEMPLATE = """ Your available action types are
"none action speak non-verbal communication leave".
Note: You can "leave" this conversation if 1. you have achieved your social goals, 2. this conversation makes you uncomfortable, 3. you find it uninteresting/you lose your patience, 4. or for other reasons you want to leave.

Please only generate a JSON string including the action type and the argument.
Your action should follow the given format:
\nAs an example, for the schema {\"properties\": {\"foo\": {\"title\": \"Foo\", \"description\": \"a list of strings\", \"type\": \"array\", \"items\": {\"type\": \"string\"}}}, \"required\": [\"foo\"]}
the object {\"foo\": [\"bar\", \"baz\"]} is a well-formatted instance of the schema. The object {\"properties\": {\"foo\": [\"bar\", \"baz\"]}} is not well-formatted.
\nHere is the output schema:\n```\n{\"description\": \"An interface for messages.\\nThere is only one required method: to_natural_language\", \"properties\": {\"action_type\": {\"title\": \"Action Type\", \"description\": \"whether to speak at this turn or choose to not do anything\", \"enum\": [\"none\", \"speak\", \"non-verbal communication\", \"action\", \"leave\"], \"type\": \"string\"}, \"argument\": {\"title\": \"Argument\", \"description\": \"the utterance if choose to speak, the expression or gesture if choose non-verbal communication, or the physical action if choose action\", \"type\": \"string\"}}, \"required\": [\"action_type\", \"argument\"]}\n```\u001b[0m
"""
# static
ACTION_LIST = "none action speak non-verbal communication leave"  # " ".join(ActionType)

ACTION_REVERSE_MAP = {"left ": "leave", "did n": "none", "said:": "speak"}

MODEL_CHECKPOINT = "meta-llama/Llama-2-13b-chat-hf"

EPISODES = jsonl_to_episodes("../../data/sotopia_pi_episodes.jsonl")

EPISODE_DICT = {ep.pk: ep for ep in EPISODES}

TOKENIZER = transformers.AutoTokenizer.from_pretrained(
    MODEL_CHECKPOINT,
    padding=False,
    truncation=False,
)


def to_natural_language(self: Any) -> str:
    match self.action_type:
        case "none":
            return "did nothing"
        case "speak":
            return f'said: "{self.argument}"'
        case "non-verbal communication":
            return f"[{self.action_type}] {self.argument}"
        case "action":
            return f"[{self.action_type}] {self.argument}"
        case "leave":
            return "left the conversation"
    return "did nothing"


SELECTED_TAG = ["gpt-4_gpt-4_v0.0.1_clean"]


def detect_action(msg: str) -> str:
    # first detect what action type is, default at none
    if msg.startswith("said:"):
        action = "speak"
    elif msg.startswith("left"):
        action = "leave"
    elif msg.startswith("[non-verbal communication]"):
        action = "non-verbal communication"
    elif msg.startswith("[action]"):
        action = "action"
    else:
        action = "none"

    return action


def generate_result(msg: str) -> str:
    action = detect_action(msg)
    result = {}
    result["action_type"] = action
    result["argument"] = ""
    # know formating argument based on action type
    match action:
        case "speak":
            # NOTE: this assume that the speech is in quotes, not ending without punctuation
            result["argument"] = msg.replace("said: ", "")[1:-1]
        case "action":
            result["argument"] = msg
        case "non-verbal communication":
            result["argument"] = msg

    str_result = json.dumps(result)

    return str_result


def surpass_max_token_check(string: str, max_token: int=MAX_TOKEN, tokenizer: transformers.AutoTokenizer=TOKENIZER) -> int:
    prompt_tokens = len(tokenizer(string)["input_ids"])
    return max(prompt_tokens - max_token, 0)


def truncate_prompt_to_length(dia_his: str, surpass_num: int, tokenizer: transformers.AutoTokenizer=TOKENIZER) -> str:
    # context_len = len(tokenizer(context)['input_ids'])
    dia_sen = dia_his.split("\n")
    remove_len = 0
    i = 0
    while remove_len < surpass_num:
        remove_len += len(tokenizer(dia_sen[i])["input_ids"])
        i += 1
    trunc_dia = "\n".join(p for p in dia_sen[i:])
    return trunc_dia


def reverse_episode_log(
    epilog: FakeEpisodeLog, later_speak: bool=False, include_format: bool=True, max_token: int=MAX_TOKEN
) -> List[Dict[str, Any]]:
    episode_msg = epilog.messages
    # per episode
    if not epilog.models:
        raise Exception("No models recorded in the episode log")

    agent_model = epilog.models[1] if not later_speak else epilog.models[2]
    promt_template = PROMPT_TEMPLATE

    if len(episode_msg) > 0:
        init_loop = episode_msg[0]
        # figure out who speak later, as we must use the 2nd player's data, else turn 0 have nothing to predict the beginning
        if later_speak:
            speaker = init_loop[-1][0]  # this would be the agent as well
            turn_div = 1
        # figure out who speak the first
        else:
            speaker = init_loop[-2][0]
            turn_div = 0

    prompt_result_instances = []
    dial_history = ""
    history = []
    for i in range(0, len(episode_msg)):
        msg = episode_msg[i]
        if (len(msg) != 4) and i < (len(episode_msg) - 1):
            continue
        turn_dic = {"model": agent_model, "env_id": epilog.environment, "agent_ids": epilog.agents}
        for tpl in msg:
            if tpl[0] == "Environment" and (tpl[1] == speaker):
                if i > 0:
                    dial_history += "\n" + tpl[2]
                else:
                    # for the first context, we don't need \n
                    context = tpl[2]
                    dial_history += context

            if tpl[0] == speaker and i % 2 == turn_div:
                history.append(f"Utterance {(i - 1) // 2} by {tpl[0]} " + tpl[2])

            if tpl[0] != "Environment" and tpl[0] != speaker and i % 2 != turn_div:
                history.append(f"Utterance {(i - 1) // 2} by {tpl[0]} " + tpl[2])

            if tpl[0] == speaker:  # if speaker is the agent, use what he said as result
                str_result = generate_result(tpl[2])
                # check if this is the end
        if i % 2 == turn_div:
            # take alternative turns as we always want to predict one agent, not both
            next_turn = i
            prompt = promt_template.format(
                agent=speaker, history=dial_history, turn_number=next_turn
            )
            over_tokens = surpass_max_token_check(prompt, max_token)
            if over_tokens > 0:
                all_dial = dial_history[len(context) :]
                trun_dial = truncate_prompt_to_length(all_dial, over_tokens)
                prompt = promt_template.format(
                    agent=speaker,
                    history=context + "\n" + trun_dial,
                    turn_number=next_turn,
                )
            if include_format:
                prompt += FORMAT_TEMPLATE
            turn_dic["prompt"] = prompt
            turn_dic["result"] = str_result
            turn_dic["history"] = list(history[1:])
            turn_dic["speaker"] = speaker
            turn_dic["episode_id"] = epilog.pk
            prompt_result_instances.append(turn_dic)

    return prompt_result_instances


def concat_episode_msg(epilog: FakeEpisodeLog) -> str:
    episode_msg = epilog.messages
    # per episode

    if len(episode_msg) > 0:
        init_loop = episode_msg[0]
        speaker = init_loop[-2][0]
    dial_history = ""

    for i in range(0, len(episode_msg)):
        msg = episode_msg[i]
        if (len(msg) != 4) and i < (len(episode_msg) - 1):
            continue
        for tpl in msg:
            if tpl[0] == "Environment" and (tpl[1] == speaker):
                if i > 0:
                    dial_history += "\n" + tpl[2]
                else:
                    # for the first context, we don't need \n
                    context = tpl[2]
                    dial_history += context

    return dial_history


def parse_prompt_to_json(episode: FakeEpisodeLog, dir: str, init_speak: bool, include_format: bool=False) -> None:
    prompt_result_instances = reverse_episode_log(episode, init_speak, include_format)

    if not os.path.exists(dir):
        os.makedirs(dir)

    for i in range(len(prompt_result_instances)):
        instance = prompt_result_instances[i]
        todump = json.dumps(instance, indent=4)
        with open(dir + "/{}-{}-{}.json".format(episode.pk, instance['speaker'], i), "w") as f:
            f.write(todump)


def run_reverse_by_pk_agent(episode_pk: str, agent_side: bool, save_dir: str) -> None:
    """
    Entry function if you want to reverse engineer given a pk, not a episode
    """
    episode = EPISODE_DICT.get(episode_pk, None)
    if not episode:
        raise Exception(f"Episode {episode_pk} not found")
    parse_prompt_to_json(episode, save_dir, agent_side, True)