SalesforceDialogStudio / code /preprocess_data_OD.py
mzozulia's picture
Upload 339 files
2ea1065
"""
Copyright (c) 2023, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: Apache License 2.0
For full license text, see the LICENSE file in the repo root or https://www.apache.org/licenses/LICENSE-2.0
"""
#!/usr/bin/env python3
#
import sys, os, pdb
import json
import shutil, errno
from tqdm import tqdm
import pandas as pd
from constant import *
class PreProcessData(object):
"""docstring for PreProcessData"""
def __init__(self):
super(PreProcessData, self).__init__()
self.data_dir = "/path/to/where/the/raw/dataset/is"
self.save_dir = "/path/to/store/the/processed/dataset/" # e.g. ./data/processed/Open-Domain
def _load_json(self, path=None):
if path is None or not os.path.exists(path):
raise IOError('File does not exist: %s' % path)
# return None
with open(path) as df:
data = json.loads(df.read())
return data
def _load_txt(self, path=None, split_tok="\n", encoding="utf-8"):
if path is None or not os.path.exists(path):
raise IOError('File does not exist: %s' % path)
with open(path, 'r', encoding=encoding) as df:
data = df.read().strip().split(split_tok)
return data
def _load_csv(self, path=None, sep="\t"):
if path is None or not os.path.exists(path):
raise IOError('File does not exist: %s' % path)
with open(path) as df:
data = pd.read_csv(df, sep=sep)
return data
def _load_jsonl(self, path=None):
if path is None or not os.path.exists(path):
raise IOError('File does not exist: %s' % path)
data = []
with open(path) as df:
for line in df.readlines():
data.append(json.loads(line))
return data
def _load_dir_json(self, dir_path=None):
if dir_path is None or not os.path.exists(dir_path): return None
total_data = [] # assume data is a list of dialogs
for filename in sorted(os.listdir(dir_path)):
if filename in ["schema.json"]: continue
if not filename.endswith(".json"): continue
file_path = os.path.join(dir_path, filename)
data = self._load_json(path=file_path)
if type(data) == list:
total_data.extend(data)
else:
total_data.append(data)
return total_data
def _load_dir_txt(self, dir_path=None, file_type="txt"):
if dir_path is None or not os.path.exists(dir_path): return None
total_data = [] # assume data is a list of dialogs
for filename in sorted(os.listdir(dir_path)):
if not filename.endswith(file_type): continue
file_path = os.path.join(dir_path, filename)
data = self._load_txt(path=file_path)
if type(data) == list:
total_data.extend(data)
else:
total_data.append(data)
return total_data
def _load_dir_tsv(self, dir_path=None, sep="\t"):
if dir_path is None or not os.path.exists(dir_path): return None
total_data = None
for filename in sorted(os.listdir(dir_path)):
file_path = os.path.join(dir_path, filename)
data = self._load_csv(path=file_path, sep=sep)
total_data = pd.concat([total_data, data], ignore_index=True)
return total_data
def _save_json(self, data, path):
with open(path, "w") as tf:
json.dump(data, tf, indent=4)
def init_dial(self, dial_idx=0, ori_dial_id=""):
dial = {
ORI_DIAL_ID: ori_dial_id,
DIAL_IDX: int(dial_idx),
ORI_DIAL_INFO: {},
LOG: [],
PROMPT: [],
}
return dial
def init_turn(self, turn_id=0, dial_hist=[]):
turn = {
TURN_ID: int(turn_id),
USR_UTT: "",
SYS_UTT: "",
DIAL_HIST: " ".join(dial_hist),
ORI_USR_ANN: {},
ORI_SYS_ANN: {},
}
return turn
def save_dial(self, data, data_name="", file_idx=0, mode="train"):
save_name = f"dialogues_{file_idx}.json"
folder_path = os.path.join(self.save_dir, data_name, mode)
if not os.path.exists(folder_path): os.makedirs(folder_path)
path = os.path.join(folder_path, save_name)
self._save_json(data, path)
def save_original_examples(self, examples, data_name):
"""
save 5 original data points just for reference and check
data would be a list of length 5, each entry is a dialog
in the form of dictionary
"""
path = os.path.join(self.save_dir, data_name, "original_examples.json")
self._save_json(examples, path)
print("original examples saved")
def save_converted_examples(self, data_name):
"""
extract the first 5 examples from the train set of the
already processed data, just for reference and check
"""
data = self._load_json(os.path.join(self.save_dir, data_name, "train/dialogues_1.json"))
examples = {key: data[key] for key in list(data.keys())[:5]}
self._save_json(examples, os.path.join(self.save_dir, data_name, "converted_examples.json"))
print("converted examples saved")
def places(self):
"""
no train/val/test split"""
data_name = "PLACES3.5"
mode = "train"
data = self._load_jsonl(os.path.join(self.data_dir, data_name, "data.jsonl"))
new_data, file_idx, dial_idx = {}, 1, 1
for dial in (data):
new_dial = self.init_dial(dial_idx=dial_idx)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
for key in dial:
if key == "conversation": continue
new_dial[ORI_DIAL_INFO][key] = dial[key]
dial_hist, multiparty = [], False
for turn_idx, utt in enumerate(dial["conversation"]):
if utt.startswith("Alice:"):
new_turn = self.init_turn(turn_id=turn_idx//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = utt.split("Alice:")[-1].strip()
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
elif utt.startswith("Bob:"):
new_turn[SYS_UTT] = utt.split("Bob:")[-1].strip()
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
new_dial[LOG].append(new_turn)
elif utt.startswith("Emilie:"):
multiparty = True
break
else:
if len(utt.split(":")[0].split()) == 1:
# might have a third speaker
raise ValueError("Unknown Speaker ... ")
else:
if not turn_idx: continue
if new_turn[SYS_UTT]:
new_turn[SYS_UTT] += " " + utt
else:
new_turn[USR_UTT] += " " + utt
dial_hist[-1] += " " + utt
if multiparty: continue
new_data[new_dial_id] = new_dial
if (dial_idx) % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
dial_idx += 1
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
print(f"finishing processing {new_dial[DIAL_IDX]} dialogs for {mode} set ...")
self.save_original_examples(data[:5], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def chitchat(self):
"""
no train/val/test split"""
data_name = "chitchat-dataset"
mode = "train"
data = self._load_json(os.path.join(self.data_dir, data_name, "chitchat_dataset/dataset.json"))
new_data, file_idx, dial_idx = {}, 1, 1
for dial_id, dial in data.items():
new_dial = self.init_dial(dial_idx=dial_idx)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
new_dial[ORI_DIAL_ID] = dial_id
for key in dial:
if key == "messages": continue
new_dial[ORI_DIAL_INFO][key] = dial[key]
dial_hist, speakers = [], []
for turn in dial["messages"]:
if turn[0]["sender"] not in speakers:
speakers.append(turn[0]["sender"])
if len(speakers) < 2: continue
# if len(speakers) != 2:
# print("This is a multi-party dialog")
# continue
for turn_idx, turn in enumerate(dial["messages"]):
if turn[0]["sender"] == speakers[0]:
new_turn = self.init_turn(turn_id=turn_idx//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = " ".join([row["text"] for row in turn])
new_turn[ORI_USR_ANN]["sender"] = turn[0]["sender"]
new_turn[ORI_USR_ANN]["timestamp"] = [row["timestamp"] for row in turn]
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
elif turn[0]["sender"] == speakers[1]:
new_turn[SYS_UTT] = " ".join([row["text"] for row in turn])
new_turn[ORI_SYS_ANN]["sender"] = turn[0]["sender"]
new_turn[ORI_SYS_ANN]["timestamp"] = [row["timestamp"] for row in turn]
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
new_dial[LOG].append(new_turn)
new_data[new_dial_id] = new_dial
if (dial_idx) % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
dial_idx += 1
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
print(f"finishing processing {new_dial[DIAL_IDX]} dialogs for {mode} set ...")
self.save_original_examples({k:data[k] for k in list(data.keys())[:5]}, data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def prosocial(self):
data_name = "Prosocial"
from datasets import load_dataset
for mode in ["train", "val", "test"]:
new_data, file_idx = {}, 1
real_name = "validation" if mode == "val" else mode
data = load_dataset("allenai/prosocial-dialog", split=real_name)
data_df = data.to_pandas()
for row_id in (range(len(data_df))):
if data_df["response_id"][row_id] == 0:
new_dial = self.init_dial(dial_idx=data_df["dialogue_id"][row_id]+1)
dial_hist = []
new_turn = self.init_turn(turn_id=data_df["response_id"][row_id]+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = data_df["context"][row_id]
new_turn[SYS_UTT] = data_df["response"][row_id]
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
for key in data_df.keys():
if key in ["context", "response"]: continue
# numpy.ndarray cannot be written into json
if type(data_df[key][row_id]) == str:
new_turn[ORI_USR_ANN][key] = data_df[key][row_id]
else:
new_turn[ORI_USR_ANN][key] = data_df[key][row_id].tolist()
new_dial[LOG].append(new_turn)
if data_df["episode_done"][row_id]:
new_dial_id = f"{data_name}--{mode}--{new_dial[DIAL_IDX]}"
new_data[new_dial_id] = new_dial
if new_dial[DIAL_IDX] % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
print(f"finishing processing {new_dial[DIAL_IDX]} dialogs for {mode} set ...")
self.save_original_examples(data[:5], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def hhrlhf(self):
"""
only use the chosen pair"""
from datasets import load_dataset
data_name = "HH-RLHF"
for mode in ["train", "test"]:
data = load_dataset("Anthropic/hh-rlhf", split=mode)
data_df = data.to_pandas()
new_data, file_idx = {}, 1
for i in (range(len(data_df))):
new_dial = self.init_dial(dial_idx=i+1)
new_dial_id = f"{data_name}--{mode}--{i+1}"
dial_hist = []
utts = data_df["chosen"][i].replace("Assistant:", "Human:").split("Human:")
for turn_idx, utt in enumerate(utts[1:]):
utt = utt.replace("\n\n", " ").strip()
if turn_idx % 2 == 0:
new_turn = self.init_turn(turn_id=turn_idx//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = utt
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
else:
new_turn[SYS_UTT] = utt
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
new_dial[LOG].append(new_turn)
new_data[new_dial_id] = new_dial
if new_dial[DIAL_IDX] % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
print(f"finishing processing {new_dial[DIAL_IDX]} dialogs for {mode} set ...")
self.save_original_examples(data[:5], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def empathetic(self):
"""
consecutive turns from the same speaker happens"""
data_name = "Empathetic"
from datasets import load_dataset
for mode in ["train", "val", "test"]:
real_name = "validation" if mode == "val" else mode
data = load_dataset("empathetic_dialogues", split=real_name)
data_df = data.to_pandas()
new_data, file_idx, dial_idx, speakers = {}, 1, 1, []
for row_id in (range(len(data_df))):
utt = data_df["utterance"][row_id].replace("_comma_", ",").strip()
if data_df["utterance_idx"][row_id] == 1:
new_dial = self.init_dial(dial_idx)
new_dial[ORI_DIAL_ID] = data_df["conv_id"][row_id]
new_dial[ORI_DIAL_INFO]["context"] = data_df["context"][row_id]
new_dial[ORI_DIAL_INFO]["selfeval"] = data_df["selfeval"][row_id]
dial_hist = []
# process the first turn
new_turn = self.init_turn(turn_id=1)
new_turn[USR_UTT] = data_df["prompt"][row_id].strip()
new_turn[SYS_UTT] = utt
new_turn[ORI_USR_ANN]["tags"] = ""
new_turn[ORI_USR_ANN]["speaker_idx"] = int(data_df["speaker_idx"][row_id+1])
new_turn[ORI_SYS_ANN]["tags"] = data_df["tags"][row_id]
new_turn[ORI_SYS_ANN]["speaker_idx"] = int(data_df["speaker_idx"][row_id])
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
# speakers.append(data_df["speaker_idx"][row_id])
# in the first turn, the first speaker's utt is in the prompt and
# utterance contains the utt from the second speaker
second_speaker_id = data_df["speaker_idx"][row_id]
new_dial[LOG].append(new_turn)
new_turn = self.init_turn(turn_id=(int(data_df["utterance_idx"][row_id])+1)//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
elif data_df["speaker_idx"][row_id] == second_speaker_id:
if not new_turn[USR_UTT]: # in this case, consecutive turns from system side happens, we add utt directly to new_dial[LOG][-1]
new_dial[LOG][-1][SYS_UTT] += " " + utt
dial_hist[-1] += " " + utt
new_turn[DIAL_HIST] = " ".join(dial_hist)
else:
new_turn[SYS_UTT] = utt
new_turn[ORI_SYS_ANN]["tags"] = data_df["tags"][row_id]
new_turn[ORI_SYS_ANN]["speaker_idx"] = int(data_df["speaker_idx"][row_id])
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
new_dial[LOG].append(new_turn)
new_turn = self.init_turn(turn_id=(int(data_df["utterance_idx"][row_id])+1)//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
else:
if not new_turn[USR_UTT]:
new_turn[USR_UTT] = utt
new_turn[ORI_USR_ANN]["tags"] = data_df["tags"][row_id]
new_turn[ORI_USR_ANN]["speaker_idx"] = int(data_df["speaker_idx"][row_id])
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
else: # in this case, consecutive turns from user side happens, we add utt directly to new_turn
new_turn[USR_UTT] += " " + utt
dial_hist[-1] += " " + utt
if row_id == len(data_df)-1 or data_df["utterance_idx"][row_id+1] == 1:
# append the rest dialog in case ends with user side
if new_turn[USR_UTT]:
new_dial[LOG].append(new_turn)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
new_data[new_dial_id] = new_dial
if dial_idx % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
dial_idx += 1
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
print(f"finishing processing {dial_idx-1} dialogs for {mode} set ...")
self.save_original_examples(data[:5], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def convai2(self):
"""
incomplete dialog included, we remove dialog with equal or less than one turn"""
from datasets import load_dataset
data_name = "ConvAI2"
mode = "train"
data = load_dataset("conv_ai_2", split=mode)
data_df = data.to_pandas()
new_data, file_idx, dial_idx = {}, 1, 1
for i in (range(len(data_df))):
new_dial = self.init_dial(dial_idx=dial_idx)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
new_dial[ORI_DIAL_ID] = data_df["dialog_id"][i]
new_dial[ORI_DIAL_INFO]["id"] = data_df["id"][i]
new_dial[ORI_DIAL_INFO]["bot_profile"] = ["".join(persona) for persona in data_df["bot_profile"][i]]
new_dial[ORI_DIAL_INFO]["user_profile"] = ["".join(persona) for persona in data_df["user_profile"][i]]
new_dial[ORI_DIAL_INFO]["eval_score"] = int(data_df["eval_score"][i])
new_dial[ORI_DIAL_INFO]["profile_match"] = int(data_df["profile_match"][i])
if len(data_df["dialog"][i]) <= 2: continue
if "Text is not given." in " ".join([turn["text"] for turn in data_df["dialog"][i]]): continue
dial_hist = []
for turn_idx, turn in enumerate(data_df["dialog"][i]):
if turn_idx % 2 == 0:
new_turn = self.init_turn(turn_id=turn_idx//2+1)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = turn["text"]
new_turn[ORI_USR_ANN]["id"] = turn["id"]
new_turn[ORI_USR_ANN]["sender"] = turn["sender"]
new_turn[ORI_USR_ANN]["sender_class"] = turn["sender_class"]
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
else:
new_turn[SYS_UTT] = turn["text"]
new_turn[ORI_SYS_ANN]["id"] = turn["id"]
new_turn[ORI_SYS_ANN]["sender"] = turn["sender"]
new_turn[ORI_SYS_ANN]["sender_class"] = turn["sender_class"]
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
new_dial[LOG].append(new_turn)
if not new_turn[SYS_UTT]:
new_dial[LOG].append(new_turn)
new_data[new_dial_id] = new_dial
if new_dial[DIAL_IDX] % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
dial_idx += 1
print(f"finishing processing {dial_idx-1} dialogs for {mode} set ...")
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
self.save_original_examples(data[:5], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def antiscam(self):
"""
0: attacker
1: agent
0 always starts conversation
1 always ends conversation
"""
data_name = "AntiScam"
data = self._load_txt(os.path.join(self.data_dir, data_name, "data/AntiScam_all.txt"), encoding='latin-1')
new_data, file_idx, dial_idx, turn_idx, dial_hist = {}, 1, 1, 1, []
mode = "train"
new_dial = self.init_dial(dial_idx=dial_idx)
new_turn = self.init_turn(turn_id=turn_idx)
for row in (data):
speaker, utt = row.split("\t")
if speaker == "0":
if new_turn[SYS_UTT]: # start a new turn
# wrap up the previous turn
new_dial[LOG].append(new_turn)
turn_idx += 1
dial_hist.append(f"<{SPEAKER1.upper()}> " + new_turn[USR_UTT])
dial_hist.append(f"<{SPEAKER2.upper()}> " + new_turn[SYS_UTT])
# start a new turn
new_turn = self.init_turn(turn_id=turn_idx)
new_turn[DIAL_HIST] = " ".join(dial_hist)
new_turn[USR_UTT] = utt.strip('\"')
else: # multiple utt from '0'
new_turn[USR_UTT] += " " + utt.strip('\"')
new_turn[USR_UTT] = new_turn[USR_UTT].strip()
elif speaker == "1":
new_turn[SYS_UTT] += " " + utt.strip('"')
new_turn[SYS_UTT] = new_turn[SYS_UTT].strip()
elif not speaker: # finish a dialog
if new_turn[SYS_UTT]: # wrap up the previous turn
new_dial[LOG].append(new_turn)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
new_data[new_dial_id] = new_dial
if dial_idx % 10000 == 0:
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
new_data = {} # reset
file_idx += 1
dial_idx += 1
turn_idx = 1
dial_hist = []
new_dial = self.init_dial(dial_idx=dial_idx)
new_turn = self.init_turn(turn_id=turn_idx)
else:
raise ValueError("Unknown speaker ... ")
if new_turn[SYS_UTT]:
new_dial[LOG].append(new_turn)
new_dial_id = f"{data_name}--{mode}--{dial_idx}"
new_data[new_dial_id] = new_dial
print(f"finishing processing {dial_idx} dialogs for {mode} set ...")
self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
if new_data: self.save_dial(new_data, data_name=data_name, file_idx=file_idx, mode=mode)
self.save_original_examples(data[:150], data_name)
self.save_converted_examples(data_name)
print("*"*10, f"finishing processing dataset {data_name}", "*"*10)
def run_all(self):
# self.places()
# self.chitchat()
# self.prosocial()
# self.hhrlhf()
# self.empathetic()
# self.convai2()
self.antiscam()
def copy_example(self):
source_dir = self.save_dir
for target_dir in [ "/home/qkun/projs/TOD-Project/Datasets/Open-Domain_PROCESSED/", "/home/qkun/projs/DialogStudio-Release/open-domain-dialogues/"]:
# target_dir = "/home/qkun/projs/TOD-Project/Datasets/Open-Domain_PROCESSED/"
# target_dir2 = "/home/qkun/projs/DialogStudio-Release/open-domain-dialogues/"
file_list = ["converted_examples.json", "original_examples.json", "readme.txt", "LICENSE"]
for dir_name in sorted(os.listdir(source_dir)):
if os.path.isfile(os.path.join(source_dir, dir_name)): continue
if not os.path.exists(os.path.join(target_dir, dir_name)): os.makedirs(os.path.join(target_dir, dir_name))
for filename in file_list:
source_path = os.path.join(source_dir, dir_name, filename)
target_path = os.path.join(target_dir, dir_name, filename)
if not os.path.exists(source_path): continue
shutil.copy(source_path, target_path)
def main():
preprocess = PreProcessData()
preprocess.run_all()
preprocess.copy_example()
if __name__ == '__main__':
main()