Update data preprocessing methods

0-eval_pretrain.py

@@ -69,18 +69,11 @@ if __name__ == "__main__":
     max_seq_len = 512
     lm_config = LMConfig()
     lm_config.max_seq_len = max_seq_len
-    # 对话是否携带历史对话（当前模型太弱，增大历史上下文，基本导致胡言乱语）
-    contain_history_chat = False
     # -----------------------------------------------------------------------------
 
     model, tokenizer = init_model(lm_config)
-
     model = model.eval()
-    # 推送到huggingface
+    # int(input('输入0自动测试，输入1问题测试：'))
-    # model.push_to_hub("minimind")
-    # tokenizer.push_to_hub("minimind")
-
-    # answer_way = int(input('输入0自动测试，输入1问题测试：'))
     answer_way = 0
     stream = True
 
@@ -101,15 +94,9 @@ if __name__ == "__main__":
         '江苏省的最好的大学',
     ]
 
-    messages_origin = []
-    messages = messages_origin
-
     qa_index = 0
     while True:
         start = time.time()
-        if not contain_history_chat:
-            messages = messages_origin.copy()
-
         if answer_way == 1:
             # run generation
             prompt = input('用户：')
@@ -120,20 +107,9 @@ if __name__ == "__main__":
             print('问题：', prompt)
             qa_index += 1
 
-        messages.append({"role": "user", "content": prompt})
-
-        # print(messages)
-        new_prompt = tokenizer.apply_chat_template(
-            messages,
-            tokenize=False,
-            add_generation_prompt=True
-        )[-(max_seq_len - 1):]
-
         x = tokenizer(prompt).data['input_ids']
         x = (torch.tensor(x, dtype=torch.long, device=device)[None, ...])
 
-        answer = new_prompt
-
         with torch.no_grad():
             res_y = model.generate(x, tokenizer.eos_token_id, max_new_tokens=max_seq_len, temperature=temperature,
                                    top_k=top_k, stream=stream)
@@ -172,8 +148,5 @@ if __name__ == "__main__":
 
             print('\n')
 
-        if contain_history_chat:
-            assistant_answer = answer.replace(new_prompt, "")
-            messages.append({"role": "assistant", "content": assistant_answer})
         end = time.time()
         print(end - start, 's')

1-pretrain.py

@@ -4,6 +4,8 @@ import argparse
 import time
 import math
 import warnings
+
+import pandas as pd
 import torch
 import torch.distributed as dist
 from torch import optim
@@ -11,6 +13,9 @@ from torch.nn.parallel import DistributedDataParallel
 from torch.optim.lr_scheduler import CosineAnnealingLR
 from torch.utils.data import DataLoader, DistributedSampler
 from contextlib import nullcontext
+
+from transformers import AutoTokenizer
+
 from model.model import Transformer
 from model.LMConfig import LMConfig
 from model.dataset import PretrainDataset
@@ -98,11 +103,13 @@ def init_model():
     def count_parameters(model):
         return sum(p.numel() for p in model.parameters() if p.requires_grad)
 
+    tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
+
     model = Transformer(lm_config).to(args.device)
-    moe_path = '_moe' if lm_config.use_moe else ''
+    # moe_path = '_moe' if lm_config.use_moe else ''
 
     Logger(f'LLM总参数量：{count_parameters(model) / 1e6:.3f} 百万')
-    return model
+    return model, tokenizer
 
 
 def init_distributed_mode():
@@ -122,14 +129,15 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="MiniMind Pretraining")
     parser.add_argument("--out_dir", type=str, default="out", help="Output directory")
     parser.add_argument("--epochs", type=int, default=20, help="Number of epochs")
-    parser.add_argument("--batch_size", type=int, default=32, help="Batch size")
+    parser.add_argument("--batch_size", type=int, default=48, help="Batch size")
     parser.add_argument("--learning_rate", type=float, default=2e-4, help="Learning rate")
-    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="Device to use")
+    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu",
+                        help="Device to use")
     parser.add_argument("--dtype", type=str, default="bfloat16", help="Data type")
     parser.add_argument("--use_wandb", action="store_true", help="Use Weights & Biases")
     parser.add_argument("--wandb_project", type=str, default="MiniMind-Pretrain", help="Weights & Biases project name")
     parser.add_argument("--num_workers", type=int, default=8, help="Number of workers for data loading")
-    parser.add_argument("--data_path", type=str, default="./dataset/pretrain_data.bin", help="Path to training data")
+    parser.add_argument("--data_path", type=str, default="./dataset/pretrain_data.csv", help="Path to training data")
     parser.add_argument("--ddp", action="store_true", help="Use DistributedDataParallel")
     parser.add_argument("--accumulation_steps", type=int, default=8, help="Gradient accumulation steps")
     parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping threshold")
@@ -160,12 +168,15 @@ if __name__ == "__main__":
 
     if args.use_wandb and (not ddp or ddp_local_rank == 0):
         import wandb
+
         wandb.init(project=args.wandb_project, name=args.wandb_run_name)
     else:
         wandb = None
 
-    data_path_list = [args.data_path]
+    model, tokenizer = init_model()
-    train_ds = PretrainDataset(data_path_list, max_length=max_seq_len, memmap=True)
+    df = pd.read_csv(args.data_path)
+    df = df.sample(frac=1.0)
+    train_ds = PretrainDataset(df, tokenizer, max_length=max_seq_len)
     train_sampler = DistributedSampler(train_ds) if ddp else None
     train_loader = DataLoader(
         train_ds,
@@ -177,8 +188,6 @@ if __name__ == "__main__":
         sampler=train_sampler
     )
 
-    model = init_model()
-
     scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
     optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
 

README.md

@@ -78,7 +78,18 @@ https://github.com/user-attachments/assets/88b98128-636e-43bc-a419-b1b1403c2055
 ### 👉**最近更新**
 
 <details close> 
-<summary> <b>2024-09-17 (new🎉)</b> </summary>
+<summary> <b>2024-09-27</b> </summary>
+
+- 09-27更新pretrain数据集的预处理方式，为了保证文本完整性，放弃预处理成.bin训练的形式（轻微牺牲训练速度）。
+
+- 目前pretrain预处理后的文件命名为：pretrain_data.csv。
+
+- 删除了一些冗余的代码。
+
+</details>
+
+<details close> 
+<summary> <b>2024-09-17</b> </summary>
 
 - 更新minimind-v1-moe模型
 

README_en.md

@@ -84,6 +84,17 @@ We hope this open-source project helps LLM beginners get started quickly!
 
 ### 👉**Recent Updates**
 
+<details close> 
+<summary> <b>2024-09-27</b> </summary>
+
+- Updated the preprocessing method for the pretrain dataset on 09-27 to ensure text integrity, opting to abandon the preprocessing into .bin training format (slightly sacrificing training speed).
+
+- The current filename for the pretrain data after preprocessing is: pretrain_data.csv.
+
+- Removed some redundant code.
+
+</details>
+
 <details close> 
 <summary> <b>2024-09-17 (new🎉)</b> </summary>
 

data_process.py

@@ -1,3 +1,4 @@
+import csv
 import itertools
 import re
 import json
@@ -12,62 +13,15 @@ from datasets import load_dataset
 bos_token = "<s>"
 eos_token = "</s>"
 
-# pretrain
-def process_wiki_clean():
-    with open('./dataset/clean-wikipedia-cn.json', 'r', encoding='utf-8') as f_read:
-        data = [ujson.loads(line) for line in f_read]
-    data_len = len(data)
-    doc_ids = []
-    for idx, line in enumerate(data):
-        text = line['response']
-        text_id = tokenizer(f'{bos_token}{text}{eos_token}').data['input_ids']
-        if len(text_id) > 5:
-            doc_ids += text_id
-        if idx % (int(data_len / 20)) == 0:
-            print(f"[{idx}/{data_len}] {text}")
-    arr = np.array(doc_ids, dtype=np.uint16)
-    with open('./dataset/clean-wikipedia-cn.bin', 'wb') as f:
-        f.write(arr.tobytes())
 
-
+def pretrain_process(chunk_size=50000):
-# pretrain
-def process_other():
-    data = []
-
-    with open('./dataset/alpaca_gpt4_data_zh.json', 'r', encoding='utf-8') as f:
-        data_ = json.load(f)
-        data += data_
-
-    with open('./dataset/alpaca_data_zh_51k.json', 'r', encoding='utf-8') as f:
-        data_ = json.load(f)
-        data += data_
-
-    doc_ids = []
-    for idx, per in enumerate(data):
-        q = per['instruction']
-        i = per['input']
-        a = per['output']
-        q = q + i
-        if len(q) < 10 or len(a) < 5:
-            continue
-        if len(q) > 256 or len(a) > 256:
-            continue
-        text_id = tokenizer(f'{bos_token}{q}，{a}{eos_token}').data['input_ids']
-        if len(text_id) > 5:
-            doc_ids += text_id
-        if idx % 50000 == 0:
-            print(idx, len(data))
-
-    arr = np.array(doc_ids, dtype=np.uint16)
-    with open('./dataset/clean_other.bin', 'wb') as f:
-        f.write(arr.tobytes())
-
-
-def process_seq_monkey(chunk_size=50000):
-    doc_ids = []
     chunk_idx = 0
 
     with jsonlines.open('./dataset/mobvoi_seq_monkey_general_open_corpus.jsonl') as reader:
+        with open('./dataset/pretrain_data.csv', 'w', newline='', encoding='utf-8') as csvfile:
+            writer = csv.writer(csvfile)
+            writer.writerow(['text'])
+
             while True:
                 chunk = list(itertools.islice(reader, chunk_size))
                 if not chunk:
@@ -78,44 +32,12 @@ def process_seq_monkey(chunk_size=50000):
                         content = obj.get('text', '')
                         if len(content) > 512:
                             continue
-                    text_id = tokenizer(f'{bos_token}{content}{eos_token}').data['input_ids']
+                        writer.writerow([content])
-                    doc_ids += text_id
                     except UnicodeDecodeError as e:
                         print(f"Skipping invalid line {chunk_idx * chunk_size + idx + 1}: {e}")
                         continue
-
                 chunk_idx += 1
-            print(f"Processed chunk {chunk_idx} with {chunk_size} lines")
+                print('chunk:', ((chunk_idx - 1) * chunk_size, chunk_idx * chunk_size), 'process end')
-
-            if len(doc_ids) > 1000000:
-                arr = np.array(doc_ids, dtype=np.uint16)
-                with open(f'./dataset/clean_seq_monkey.bin', 'ab') as f:
-                    f.write(arr.tobytes())
-                doc_ids = []
-
-    if doc_ids:
-        arr = np.array(doc_ids, dtype=np.uint16)
-        with open(f'./dataset/clean_seq_monkey.bin', 'ab') as f:
-            f.write(arr.tobytes())
-
-
-def pretrain_process():
-    # process_wiki_clean()
-    process_seq_monkey()
-
-    data_path_list = [
-        # './dataset/clean-wikipedia-cn.bin',
-        './dataset/clean_seq_monkey.bin'
-    ]
-    data_lst = []
-    for data_path in data_path_list:
-        with open(data_path, 'rb') as f:
-            data = np.fromfile(f, dtype=np.uint16)
-            data_lst.append(data)
-    arr = np.concatenate(data_lst)
-    print(arr.shape)
-    with open('./dataset/pretrain_data.bin', 'wb') as f:
-        f.write(arr.tobytes())
 
 
 def sft_process(contain_history=False):
@@ -186,6 +108,7 @@ def sft_process(contain_history=False):
                 process_and_write_data(data)
                 data = []
 
+
 def rl_process():
     ################
     # Dataset

model/dataset.py

@@ -13,37 +13,28 @@ os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
 
 class PretrainDataset(Dataset):
-    def __init__(self, data_path_lst, max_length=512, memmap=False):
+    def __init__(self, df, tokenizer, max_length=512):
         super().__init__()
-        #
+        self.df = df
-        if memmap:
+        self.tokenizer = tokenizer
-            with open(data_path_lst[0], 'r') as f:
+        self.max_length = max_length
-                nbytes = f.seek(0, 2)
+        self.padding = 0
-                flen = f.tell() // np.dtype('uint16').itemsize
-            self.data = np.memmap(data_path_lst[0], dtype=np.dtype('uint16'), shape=(flen // max_length, max_length))
-        else:
-            data_lst = []
-            for data_path in data_path_lst:
-                with open(data_path, 'rb') as f:
-                    data = np.fromfile(f, dtype=np.uint16)
-                    data_lst.append(data)
-            data = np.concatenate(data_lst)
-            data = data[:max_length * int(len(data) / max_length)]
-            # np.random.shuffle(data)
-            self.data = data.reshape(-1, max_length)
-        #
-        print("memmap:{} train data.shape:{}".format(memmap, self.data.shape))
-        print("downloading finished.....")
 
     def __len__(self):
-        return self.data.shape[0]
+        return self.df.shape[0]
 
     def __getitem__(self, index: int):
         #
-        sample = self.data[index]
+        sample = self.df.iloc[index]
-        X = np.array(sample[:-1]).astype(np.int64)
+        text = f"{self.tokenizer.bos_token}{str(sample['text'])}{self.tokenizer.eos_token}"
-        Y = np.array(sample[1:]).astype(np.int64)
+        input_id = self.tokenizer(text).data['input_ids'][:self.max_length]
+        # 没满最大长度的剩余部分
+        padding_len = self.max_length - len(input_id)
+        input_id = input_id + [self.padding] * padding_len
 
+        input_id = np.array(input_id)
+        X = np.array(input_id[:-1]).astype(np.int64)
+        Y = np.array(input_id[1:]).astype(np.int64)
         return torch.from_numpy(X), torch.from_numpy(Y)
 
 
@@ -56,7 +47,7 @@ class SFTDataset(Dataset):
         self.answer_max_len = answer_max_len
         #
         self.tokenizer = tokenizer
-        self.padding = 0  # self.tokenizer.special_tokens['<pad>']
+        self.padding = 0
         self.bos_id = self.tokenizer('<s>assistant').data['input_ids']
 
     def __len__(self):