cfg3f_dataset

CFG3f Dataset Collection

This repository contains three versions of the CFG3f synthetic dataset, generated using Context-Free Grammar rules for language modeling research.

📊 Dataset Overview

🔧 Dataset Structure

cfg3f_datasets/
├── balanced/           # Balanced generation (all rules equal probability)
│   ├── train/         # 800K training samples
│   └── val/           # 80K validation samples
├── moderate/          # Moderate imbalance (10:1 ratio)
│   ├── train/         # 800K training samples  
│   └── val/           # 80K validation samples
└── extreme/           # Extreme imbalance (100:1 ratio)
    ├── train/         # 800K training samples
    └── val/           # 80K validation samples

🎯 Grammar Rules (CFG3f)

The dataset is generated using a hierarchical context-free grammar with 4 levels:

Level 4 (Highest Abstraction)

22 -> 20 21 | 20 19 21 | 21 19 19 | 20 20
21 -> 18 17 | 17 16 | 16 17 18 | 16 18
20 -> 16 17 | 17 16 18
19 -> 18 16 18 | 17 18 | 18 18

Level 3

18 -> 14 15 13 | 15 13 13 | 13 15
17 -> 14 15 | 15 14 | 15 14 13
16 -> 13 15 13 | 15 15 | 14 13 | 14 14

Level 2

15 -> 10 11 11 | 11 11 10 | 10 10 | 12 12 11
14 -> 12 10 12 | 10 12 12 | 12 11 | 10 12
13 -> 11 12 | 10 12 11 | 12 11 12

Level 1

12 -> 7 9 7 | 9 8 | 8 8 9
11 -> 8 8 | 9 7 | 9 7 7
10 -> 8 9 9 | 9 7 9 | 7 9 9

Level 0 (Terminals)

9 -> '3' '3' | '1' '1' | '1' '2' '1'
8 -> '3' '3' '1' | '3' '1' '1' | '1' '2'
7 -> '3' '2' | '3' '2' '2' | '3' '1' '2' | '2' '2' '1'

⚖️ Weight Configurations

Balanced Version

• All grammar rules have equal probability
• Explores all possible generation paths uniformly
• Results in longer, more diverse sentences

Moderate Version (10:1 ratio)

MODERATE_WEIGHTS = {
    '9': [1, 10, 1],   # Prefer "1 1" over others
    '8': [1, 1, 10],   # Prefer "1 2" over others
    '7': [10, 1, 1, 1], # Prefer "3 2" over others
    # ... similar patterns for all rules
}

Extreme Version (100:1 ratio)

EXTREME_WEIGHTS = {
    '9': [1, 100, 1],   # Heavily prefer "1 1"
    '8': [1, 1, 100],   # Heavily prefer "1 2"
    '7': [100, 1, 1, 1], # Heavily prefer "3 2"
    # ... similar patterns for all rules
}

🔀 Data Format

• File format: Binary (.bin) files compatible with NanoGPT
• Token encoding:
  • '1' → 0
  • '2' → 1
  • '3' → 2
  • '[BOS]' → 3
  • '[EOS]' → 4
• Header: 256 int32 values followed by uint16 tokens
• Vocabulary size: 7 tokens (including special tokens)

📈 Key Differences

🚀 Usage Example

import numpy as np

def read_bin_file(filename):
    with open(filename, 'rb') as f:
        # Read header
        header = np.frombuffer(f.read(256 * 4), dtype=np.int32)
        magic, version, num_tokens = header[0], header[1], header[2]
        
        # Read tokens
        tokens = np.frombuffer(f.read(), dtype=np.uint16)
        return tokens

# Load training data
train_tokens = read_bin_file("balanced/train/cfg3f_train_000000.bin")
print(f"Loaded {len(train_tokens)} tokens")

🎓 Research Applications

• Long-tail Learning: Study how models handle imbalanced distributions
• Grammar Induction: Test models' ability to learn hierarchical structure
• Generalization: Evaluate performance on rare vs common patterns
• Synthetic Benchmarks: Controlled experiments with known ground truth

📝 Citation

If you use this dataset in your research, please cite:

@dataset{cfg3f_dataset_2024,
  title={CFG3f: Context-Free Grammar Dataset with Controlled Imbalance},
  author={Shuche Wang},
  year={2024},
  url={https://huggingface.co/datasets/Shuche/cfg3f_dataset}
}

🔗 Related Work

This dataset is part of research on:

• Sharpness-aware optimization
• Imbalanced learning in language models
• Synthetic data generation for NLP

📄 License

MIT License - Feel free to use for research and educational purposes.