|
|
import json |
|
|
import os |
|
|
import re |
|
|
from copy import deepcopy |
|
|
from typing import Any, Dict, List |
|
|
|
|
|
import click |
|
|
from db_free_reverse_engineering import run_reverse_by_pk_agent |
|
|
from tqdm import tqdm |
|
|
|
|
|
|
|
|
def get_attributed_data(data: List[Dict[str, Any]], utterance_pattern: str) -> List[Dict[str, Any]]: |
|
|
attributed_data = [] |
|
|
print(f"Processing {len(data)} episodes") |
|
|
for d in tqdm(data): |
|
|
for uttr_key, attributed_uttr in d['attributed_utterances'].items(): |
|
|
match = re.search(utterance_pattern, uttr_key) |
|
|
if match: |
|
|
turn_number = match.group(1) |
|
|
agent_name = match.group(2) |
|
|
else: |
|
|
raise Exception(f"Utterance key not in correct format: {uttr_key}") |
|
|
if agent_name != d['agent']: |
|
|
continue |
|
|
|
|
|
utterance_path = f"../../data/episode_utterances/{d['episode_id']}-{d['agent']}-{turn_number}.json" |
|
|
if not os.path.exists(utterance_path): |
|
|
raise Exception(f"Utterance not found: {utterance_path}") |
|
|
|
|
|
with open(f"../../data/episode_utterances/{d['episode_id']}-{d['agent']}-{turn_number}.json", 'r') as f: |
|
|
sotopia_utterance = json.load(f) |
|
|
|
|
|
new_utterance = deepcopy(sotopia_utterance) |
|
|
new_utterance['attribution'] = attributed_uttr[1] |
|
|
new_utterance['turn_number'] = turn_number |
|
|
new_utterance['goal_score'] = d['goal_score'] |
|
|
|
|
|
attributed_data.append(new_utterance) |
|
|
return attributed_data |
|
|
|
|
|
|
|
|
@click.command() |
|
|
@click.option("--data_dir", type=str, required=True, help="Directory containing data files.") |
|
|
@click.option("--input_file", type=str, required=True, help="Path to the raw JSON file.") |
|
|
@click.option("--reward_output_file", type=str, required=True, help="Path to the processed JSON file.") |
|
|
@click.option("--sft_output_file", type=str, required=True, help="Path to the processed JSON file.") |
|
|
def main(data_dir: str, input_file: str, reward_output_file: str, sft_output_file: str) -> None: |
|
|
with open(os.path.join(data_dir, input_file), 'r') as f: |
|
|
data: List[Dict[str, Any]] = [json.loads(d) for d in f.readlines()] |
|
|
|
|
|
cache_dir = os.path.join(data_dir, "episode_utterances") |
|
|
if not os.path.exists(cache_dir): |
|
|
os.makedirs(cache_dir) |
|
|
for d in tqdm(data): |
|
|
run_reverse_by_pk_agent(d['episode_id'], True, cache_dir) |
|
|
run_reverse_by_pk_agent(d['episode_id'], False, cache_dir) |
|
|
|
|
|
utterance_pattern = r'Utterance (\d+) by ([A-Za-z ]+)' |
|
|
print("turning into attributed utterances") |
|
|
|
|
|
attributed_data = get_attributed_data(data, utterance_pattern) |
|
|
|
|
|
def calc_reward(utter_attrib: float, goal_score: float) -> float: |
|
|
if utter_attrib == -1: |
|
|
reward = -1.0 |
|
|
else: |
|
|
reward = utter_attrib / 3 * goal_score |
|
|
return reward |
|
|
|
|
|
sotopia_pi_utterance_reward = [] |
|
|
for d in tqdm(attributed_data): |
|
|
sotopia_pi_utterance_reward.append( |
|
|
{ |
|
|
"instruction": d['prompt'], |
|
|
"output": d['result'], |
|
|
"value": calc_reward(d['attribution'], d['goal_score']), |
|
|
} |
|
|
) |
|
|
|
|
|
with open(os.path.join(data_dir, reward_output_file), 'w') as f: |
|
|
json.dump(sotopia_pi_utterance_reward, f, indent=4) |
|
|
|
|
|
sotopia_pi_utterance_sft = [] |
|
|
for d in tqdm(attributed_data): |
|
|
sotopia_pi_utterance_sft.append( |
|
|
{ |
|
|
"instruction": d['prompt'], |
|
|
"input": "", |
|
|
"output": d["result"], |
|
|
} |
|
|
) |
|
|
|
|
|
with open(os.path.join(data_dir, sft_output_file), 'w') as f: |
|
|
json.dump(sotopia_pi_utterance_sft, f, indent=4) |
|
|
|
|
|
if __name__ == "__main__": |
|
|
main() |
|
|
|
