Mini-LLM / data /README.md
Ashx098's picture
Upload folder using huggingface_hub
f4e346e verified
|
Raw
History Blame Contribute Delete
6.67 kB

Data Module

This module handles all data preprocessing, tokenization, and preparation for training.

Overview

The data pipeline converts raw text into binary token files optimized for training:

  • Raw text collection from multiple sources
  • Tokenization using BPE tokenizer
  • Binary serialization for efficient loading
  • Train/validation splitting

Directory Structure

data/
β”œβ”€β”€ raw/                    # Raw text sources
β”‚   β”œβ”€β”€ books/             # Book corpus
β”‚   β”œβ”€β”€ wikipedia/         # Wikipedia dumps
β”‚   β”œβ”€β”€ fineweb/           # Web crawl data
β”‚   └── merged_text/
β”‚       └── corpus.txt     # Combined corpus
β”œβ”€β”€ bin/                   # Tokenized binary files
β”‚   β”œβ”€β”€ train.bin         # Training data (uint16)
β”‚   └── val.bin           # Validation data (uint16)
└── prepare_data.py       # Tokenization script

Data Processing Pipeline

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ 1. Raw Text Sources                         β”‚
β”‚    - Books: 15 files                        β”‚
β”‚    - Wikipedia: 3 dumps                     β”‚
β”‚    - FineWeb: 1 crawl                       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                   β”‚
                   β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ 2. Merge & Clean                            β”‚
β”‚    β†’ corpus.txt (all text combined)         β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                   β”‚
                   β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ 3. Tokenize (prepare_data.py)              β”‚
β”‚    - Load BPE tokenizer                     β”‚
β”‚    - Process line-by-line                   β”‚
β”‚    - Append EOS tokens                      β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                   β”‚
                   β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ 4. Convert to NumPy (uint16)               β”‚
β”‚    - Vocab size: 32,000 fits in uint16     β”‚
β”‚    - Memory efficient (2 bytes/token)       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                   β”‚
                   β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ 5. Train/Val Split (90/10)                 β”‚
β”‚    - train.bin: 325M tokens                 β”‚
β”‚    - val.bin: 36M tokens                    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Data Preparation Script

File: prepare_data.py

import numpy as np
from transformers import AutoTokenizer
from tqdm import tqdm

# 1. Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Tokenizer/BPE")
eos_id = tokenizer.eos_token_id

# 2. Read corpus
with open("data/raw/merged_text/corpus.txt") as f:
    lines = f.readlines()

# 3. Tokenize
all_tokens = []
for line in tqdm(lines):
    tokens = tokenizer.encode(line.strip())
    tokens.append(eos_id)  # Mark end of line
    all_tokens.extend(tokens)

# 4. Convert to uint16
ids = np.array(all_tokens, dtype=np.uint16)

# 5. Split
val_count = int(len(ids) * 0.1)
train_ids = ids[:-val_count]
val_ids = ids[-val_count:]

# 6. Save
train_ids.tofile("data/bin/train.bin")
val_ids.tofile("data/bin/val.bin")

Example: Text β†’ Tokens

Input Text (corpus.txt):

The quick brown fox jumps over the lazy dog.
Machine learning is transforming the world.

Tokenization Process:

Line 1: "The quick brown fox jumps over the lazy dog."
  Tokens: [1, 334, 3855, 288, 267, 2959, 354, 267, 12397, 8885, 2]
          [<s>, The, quick, brown, fox, jumps, over, the, lazy, dog, </s>]

Line 2: "Machine learning is transforming the world."
  Tokens: [1, 5234, 1234, 456, 7890, 267, 9876, 2]
          [<s>, Machine, learning, is, transforming, the, world, </s>]

Combined: [1, 334, 3855, ..., 2, 1, 5234, ..., 2]

Binary Format:

train.bin structure:
  Byte 0-1:   Token 0 (uint16)
  Byte 2-3:   Token 1 (uint16)
  Byte 4-5:   Token 2 (uint16)
  ...
  Byte N-2:N  Token N/2 (uint16)

Total size: 325,004,796 tokens Γ— 2 bytes = ~650 MB

Dataset Statistics

Corpus Size

Raw Text:
  - Total files: 19
  - Total size: ~1.4 GB
  - Total lines: ~5.2M

Tokenized:
  - Total tokens: 361,116,440
  - Train tokens: 325,004,796 (90%)
  - Val tokens: 36,111,644 (10%)

Usage

Prepare Data

# Tokenize corpus
python data/prepare_data.py

Output:

Loading tokenizer from Tokenizer/BPE...
Vocab size: 32000
EOS ID: 2
Reading data/raw/merged_text/corpus.txt...
Total lines: 5,234,567
Tokenizing...
100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 5.2M/5.2M [02:34<00:00]
Total tokens: 361,116,440
Train tokens: 325,004,796
Val tokens:   36,111,644
βœ… Saved binary files to data/bin/

Load in Training

from train.dataloader import DataLoader

loader = DataLoader("data/bin", batch_size=16, block_size=512, split="train")
x, y = loader.get_batch(device="cuda")

# x: [16, 512] input tokens
# y: [16, 512] target tokens (shifted by 1)

Memory-Mapped Loading

The binary files are loaded using np.memmap for efficiency:

# Traditional loading (BAD)
data = np.fromfile("train.bin", dtype=np.uint16)  # Loads 650MB into RAM!

# Memory-mapped loading (GOOD)
data = np.memmap("train.bin", dtype=np.uint16, mode='r')  # OS handles paging

Benefits:

  • No RAM overhead: File stays on disk
  • Fast random access: OS caches hot pages
  • Scalable: Works with TB-scale datasets

References