import torch from torch.utils.data import Dataset from tqdm import tqdm import copy import numpy as np import pdb import os import io import json import gzip import zstandard as zstd import numpy as np from tqdm import tqdm import pandas as pd class SemiNATForSingleRoundMaskInput(Dataset): ''' Mask掉了所有的输入，只有输出的loss ''' def __init__(self, tokenizer, datas, max_length, proc): self.tokenizer = tokenizer self.max_length = max_length self.proc = proc # 用 apply + 并行加速预处理 processed = self._vectorized_preprocess(datas) self.input_ids = processed["input_ids"] self.labels = processed["labels"] self.attention_mask = processed["attention_mask"] self.slice_indices = processed["slice_indices"] def _vectorized_preprocess(self, datas): # 批量预分配内存 input_ids = np.zeros((len(datas), self.max_length), dtype=np.int64) attention_mask = np.zeros((len(datas), self.max_length), dtype=np.int64) labels = np.full((len(datas), self.max_length), -100, dtype=np.int64) slice_indices = np.full((len(datas), self.max_length), -1, dtype=np.int64) # 批量处理所有行的 messages def process_row(row): total_inputs = [] total_labels = [] sample_slice = [] # pdb.set_trace() for msg in row['messages']: # 批量分词（假设 msg['content'] 是文本列表） inputs = self.tokenizer(msg['content'], padding=False, truncation=False, add_special_tokens=False).input_ids total_inputs.extend(inputs) if msg['role'] == 'user': total_labels.extend(len(inputs) * [-100]) elif msg['role'] == 'assistant': total_labels.extend(inputs) sample_slice.extend(msg['split_pos']) # 截断或填充逻辑 seq_len = min(len(total_inputs), self.max_length) # 输入和标签 input_ids = total_inputs[:self.max_length] + [ self.tokenizer.pad_token_id ] * (self.max_length - seq_len) labels = total_labels[:self.max_length] + [-100] * ( self.max_length - seq_len) if all(l == -100 for l in labels): return None # 这一条数据无有效标签，丢弃 # attention_mask attention_mask = [1] * seq_len + [0] * (self.max_length - seq_len) # slice_indices slice_arr = np.array(sample_slice[:self.max_length] + [-1] * (self.max_length - len(sample_slice))) slice_arr[slice_arr >= self.max_length - 1] = -1 # 过滤超长位置,这里-1是因为max length是1024，最大index是1023 return input_ids, labels, attention_mask, slice_arr # 并行处理所有行（需安装 pandarallel） try: from pandarallel import pandarallel pandarallel.initialize(nb_workers=self.proc, progress_bar=True) processed = datas.parallel_apply(process_row, axis=1) except ImportError: processed = datas.progress_apply(process_row, axis=1) # tqdm 进度条 processed = processed[processed.notnull()].reset_index(drop=True) # 合并结果 # pdb.set_trace() for idx, (i_ids, lbl, attn, slc) in enumerate(processed): input_ids[idx] = i_ids labels[idx] = lbl attention_mask[idx] = attn slice_indices[idx] = slc # return { # "input_ids": input_ids, # "labels": labels, # "attention_mask": attention_mask, # "slice_indices": slice_indices # } processed_len = len(processed) return { "input_ids": input_ids[:processed_len], "labels": labels[:processed_len], "attention_mask": attention_mask[:processed_len], "slice_indices": slice_indices[:processed_len] } def __len__(self): return len(self.input_ids) def __getitem__(self, index): # 直接返回预分配的张量，避免重复转换 return (torch.as_tensor(self.input_ids[index]), torch.as_tensor(self.labels[index]), torch.as_tensor(self.attention_mask[index]), torch.as_tensor(self.slice_indices[index])) class SemiNATForSingleRound(Dataset): def __init__(self, tokenizer, datas, max_length, proc): self.tokenizer = tokenizer self.max_length = max_length self.proc = proc # 用 apply + 并行加速预处理 processed = self._vectorized_preprocess(datas) self.input_ids = processed["input_ids"] self.labels = processed["labels"] self.attention_mask = processed["attention_mask"] self.slice_indices = processed["slice_indices"] def _vectorized_preprocess(self, datas): # 批量预分配内存 input_ids = np.zeros((len(datas), self.max_length), dtype=np.int64) attention_mask = np.zeros((len(datas), self.max_length), dtype=np.int64) labels = np.full((len(datas), self.max_length), -100, dtype=np.int64) slice_indices = np.full((len(datas), self.max_length), -1, dtype=np.int64) # 批量处理所有行的 messages def process_row(row): total_inputs = [] sample_slice = [] for msg in row['messages']: # 批量分词（假设 msg['content'] 是文本列表） inputs = self.tokenizer(msg['content'], padding=False, truncation=False, add_special_tokens=False).input_ids total_inputs.extend(inputs) # 直接使用列表扩展 slice sample_slice.extend(msg['split_pos']) # 截断或填充逻辑 seq_len = min(len(total_inputs), self.max_length) # 输入和标签 input_ids = total_inputs[:self.max_length] + [ self.tokenizer.pad_token_id ] * (self.max_length - seq_len) labels = total_inputs[:self.max_length] + [-100] * ( self.max_length - seq_len) # attention_mask attention_mask = [1] * seq_len + [0] * (self.max_length - seq_len) # slice_indices slice_arr = np.array(sample_slice[:self.max_length] + [-1] * (self.max_length - len(sample_slice))) slice_arr[slice_arr > self.max_length] = -1 # 过滤超长位置 return input_ids, labels, attention_mask, slice_arr # 并行处理所有行（需安装 pandarallel） try: from pandarallel import pandarallel pandarallel.initialize(nb_workers=self.proc, progress_bar=True) processed = datas.parallel_apply(process_row, axis=1) except: processed = datas.progress_apply(process_row, axis=1) # tqdm 进度条 # 合并结果 for idx, (i_ids, lbl, attn, slc) in enumerate(processed): input_ids[idx] = i_ids labels[idx] = lbl attention_mask[idx] = attn slice_indices[idx] = slc return { "input_ids": input_ids, "labels": labels, "attention_mask": attention_mask, "slice_indices": slice_indices } def __len__(self): return len(self.input_ids) def __getitem__(self, index): # 直接返回预分配的张量，避免重复转换 return (torch.as_tensor(self.input_ids[index]), torch.as_tensor(self.labels[index]), torch.as_tensor(self.attention_mask[index]), torch.as_tensor(self.slice_indices[index])) class SemiNATDatasetForPretrain(Dataset): # data is jsonl.zstd or json.gz file def __init__(self, tokenizer, data_files, max_length, proc, cache_path=None): if cache_path and os.path.exists(cache_path): print(f"[INFO] Loading cached data from {cache_path}") cached = torch.load(cache_path) self.input_ids = cached["input_ids"] self.labels = cached["labels"] self.attention_mask = cached["attention_mask"] self.slice_indices = cached["slice_indices"] return data = [] for filename in data_files: if filename.endswith('.zstd'): data.append( pd.DataFrame([ json.loads(line) for line in self._decompress_zst_to_string(filename).splitlines() ])) else: # json.gz file, each line a json with gzip.open(filename, 'rt', encoding='utf-8') as f: data.append(pd.DataFrame([json.loads(line) for line in f])) data = pd.concat(data, ignore_index=True) self.tokenizer = tokenizer self.max_length = max_length self.proc = proc processed = self._vectorized_preprocess(data) self.input_ids = processed["input_ids"] self.labels = processed["labels"] self.attention_mask = processed["attention_mask"] self.slice_indices = processed["slice_indices"] if type(self.input_ids) != torch.Tensor: self.input_ids = torch.tensor(self.input_ids, dtype=torch.long) self.labels = torch.tensor(self.labels, dtype=torch.long) self.attention_mask = torch.tensor(self.attention_mask, dtype=torch.long) self.slice_indices = torch.tensor(self.slice_indices, dtype=torch.long) def _decompress_zst_to_string(self, input_file): with open(input_file, 'rb') as f: dctx = zstd.ZstdDecompressor() with dctx.stream_reader(f) as reader: text_stream = io.TextIOWrapper(reader, encoding='utf-8') return text_stream.read() # 读取为字符串 def _vectorized_preprocess(self, data): input_ids = np.zeros((len(data), self.max_length), dtype=np.int64) attention_mask = np.zeros((len(data), self.max_length), dtype=np.int64) labels = np.full((len(data), self.max_length), -100, dtype=np.int64) slice_indices = np.full((len(data), self.max_length), -1, dtype=np.int64) def process_row(row): inputs = self.tokenizer(row['text'], padding=False, truncation=False, add_special_tokens=False).input_ids # slice to 8-token segments. that is, sample_slice is [1, 9, 17, 25, ...] sample_slice = (np.arange(0, len(inputs), 8) + 1).tolist() # add the end if len(inputs) % 8 != 1: sample_slice.append(len(inputs)) # 截断或填充逻辑 seq_len = min(len(inputs), self.max_length) # 输入和标签 input_ids = inputs[:self.max_length] + [ self.tokenizer.pad_token_id ] * (self.max_length - seq_len) labels = [ 50279 # ] + inputs[:self.max_length - 1] + [-100] * (self.max_length - 1 - seq_len) # attention_mask attention_mask = [1] * seq_len + [0] * (self.max_length - seq_len) # slice_indices slice_arr = np.array(sample_slice[:self.max_length] + [-1] * (self.max_length - len(sample_slice))) slice_arr[slice_arr > self.max_length] = -1 # 过滤超长位置 return input_ids, labels, attention_mask, slice_arr try: from pandarallel import pandarallel pandarallel.initialize(nb_workers=self.proc, progress_bar=True) processed = data.parallel_apply(process_row, axis=1) except ImportError: processed = data.progress_apply(process_row, axis=1) # tqdm 进度条 # 合并结果 for idx, (i_ids, lbl, attn, slc) in enumerate(processed): input_ids[idx] = i_ids labels[idx] = lbl attention_mask[idx] = attn slice_indices[idx] = slc return { "input_ids": input_ids, "labels": labels, "attention_mask": attention_mask, "slice_indices": slice_indices } def __len__(self): return len(self.input_ids) def __getitem__(self, index): return (self.input_ids[index], self.labels[index], self.attention_mask[index], self.slice_indices[index])