Beyond Language Models: Byte Models are Digital World Simulators

This repository contains the weights for the bGPT model as described in the paper Beyond Language Models: Byte Models are Digital World Simulators.

bGPT supports generative modelling via next byte prediction on any type of data and can perform any task executable on a computer, showcasing the capability to simulate all activities within the digital world, with its potential only limited by computational resources and our imagination.

You can check out the demo page, which includes examples generated by the bGPT model.

The code for bGPT is available on GitHub .

Model Description

Traditional deep learning often overlooks bytes, the basic units of the digital world, where all forms of information and operations are encoded and manipulated in binary format. Inspired by the success of next token prediction in natural language processing, we introduce bGPT, a model with next byte prediction to simulate the digital world. bGPT matches specialized models in performance across various modalities, including text, audio, and images, and offers new possibilities for predicting, simulating, and diagnosing algorithm or hardware behaviour. It has almost flawlessly replicated the process of converting symbolic music data, achieving a low error rate of 0.0011 bits per byte in converting ABC notation to MIDI format. In addition, bGPT demonstrates exceptional capabilities in simulating CPU behaviour, with an accuracy exceeding 99.99% in executing various operations. Leveraging next byte prediction, models like bGPT can directly learn from vast binary data, effectively simulating the intricate patterns of the digital world.

We provide five weights of bGPT on Hugging Face corresponding to each dataset used for pre-training:

weights-conversion.pth: bGPT pre-trained on IrishMAN for data conversion (between .abc and .mid).
weights-cpu.pth: bGPT pre-trained on CPU states for CPU state modelling (.bin).
weights-text.pth: bGPT pre-trained on Wikipedia for text generation/classification (.txt).
weights-image.pth: bGPT pre-trained on ImageNet for image generation/classification (.bmp).
weights-audio.pth: bGPT pre-trained on Librispeech for audio generation/classification (.wav).

The core components of bGPT include a 12-layer patch-level decoder, a 3-layer byte-level decoder, with a hidden size of 768, totaling 110 million parameters.

Installation

To set up the bGPT environment and install the necessary dependencies, follow these steps:

Create and Activate Conda Environment

conda create --name bgpt python=3.7.9
conda activate bgpt

Install Dependencies
```
pip install -r requirements.txt
```

Install Pytorch

pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

Download Pre-trained bGPT Weights (Optional)

For those interested in starting with pre-trained models, bGPT weights are available on Hugging Face. This step is optional but recommended for users looking to leverage the model's capabilities without training from scratch.

Usage

config.py: Configuration settings for training and inference.
cpu-simulation.py: Simulate CPU states and operations.
inference.py: Perform inference tasks (e.g., generation and conversion) using pre-trained models.
train-cls.py: Training script for classification models.
train-gen.py: Training script for generative models.
utils.py: Utility functions supporting model operations and data processing.

Configuration

The config.py file contains critical settings for training and inference, allowing for flexibility across different tasks and modalities. Here's a breakdown of the main configurations:

Training Configuration

TRAIN_FOLDERS: Specify the dataset folders for training. Multiple folders can be included.
EVAL_FOLDERS: Specify evaluation dataset folders.
PRETRAINED_PATH: Path to pre-trained weights for transfer learning and fine-tuning.
WEIGHTS_PATH & LOGS_PATH: Define locations to save trained weights and logs, respectively.
NUM_EPOCHS, LEARNING_RATE, BATCH_SIZE: Control training duration, learning rate, and batch size for optimal learning.
ACCUMULATION_STEPS: Set accumulation steps to emulate larger batch sizes, managing memory usage efficiently.
PATCH_SAMPLING_BATCH_SIZE: Adjust batch size for patch sampling during training to reduce computational load, with 0 for full batch processing.

Inference Configuration

INFERENCE_WEIGHTS_PATH: Path to weights for inference.
MODE: Determines operation mode (convert or generate), guiding the model for specific outcomes.
NUM_SAMPLES, TOP_K, TOP_P, TEMPERATURE: Set sampling strategy during inference to control the diversity of outputs.

Generative Modelling

Generative modelling with bGPT is a flexible and powerful approach to learning and generating new data across various formats. bGPT segments byte sequences into patches, predicts next patch features with a patch-level decoder, and reconstructs bytes within patches using these features with a byte-level decoder. Here's how to get started:

Prepare Your Data: bGPT can handle any computer file type, including text, images, audio, executables, and encrypted or proprietary formats, without needing specific adjustments for each modality. This capability allows for straightforward and versatile training on a wide array of data. The only thing you need to do here is simply to split your data for training and evaluation.
Adjust Configuration Settings: Modify the config.py file to tailor the training process to your needs. At a minimum, you should update the TRAIN_FOLDERS and EVAL_FOLDERS to point to your actual data directories. Also, specify where to save the trained model weights and logs by setting WEIGHTS_PATH and LOGS_PATH. You may adjust other parameters based on your specific requirements. For instance, with the default PATCH_SIZE=16 and PATCH_LENGTH=512, bGPT can model byte sequences up to 8KB. If your training files are larger, and you have sufficient computational resources, consider increasing these parameters to accommodate the larger file sizes.
Leverage Pre-trained Weights (Optional): If you wish to fine-tune a pre-trained bGPT model, set PRETRAINED_PATH to the location of the pre-trained weights and ensure LOAD_FROM_PRETRAINED=True. To train a model from scratch, simply set LOAD_FROM_PRETRAINED=False.
Start Training: Run train-gen.py to begin the training process. The script will use the configurations set in config.py and apply the training data to learn generative models capable of producing new, unseen outputs in the format of your training data.

Data Conversion

The conversion mode in bGPT adds a specialized functionality for transforming data from one format to another, leveraging the model's understanding of byte sequences across different file types. This mode supports both unidirectional and bidirectional conversions, enabling a wide range of data transformation tasks. Here's how to utilize the conversion mode effectively:

Define Conversion Mode: In your config.py file, you'll define the CONVERSION_MODE setting, which governs how files are transformed. This setting offers two options: unidirectional and bidirectional conversion, denoted by "->" and "&" respectively.
- Unidirectional Conversion: Denoted by "->", this mode signifies a one-way transformation from one format to another. For instance, if you want to convert text files to HTML, you'd set CONVERSION_MODE = "txt->html". This means the model will learn to convert text files specifically into HTML format, but not vice versa.
- Bidirectional Conversion: Denoted by "&", this mode implies a two-way transformation between formats. For example, setting CONVERSION_MODE = "wav&mp3" instructs the model to learn bidirectional conversion between WAV and MP3 audio formats. In this mode, the model learns to convert files from WAV to MP3 and vice versa, allowing flexibility in both directions of conversion.
Prepare Your Data: Ensure your data pairs are stored within the same directory path in both TRAIN_FOLDERS and EVAL_FOLDERS. Each pair should share identical paths, including filenames, differing only in their file extensions. For instance, if converting between WAV and MP3 formats, ensure files like "path/audio.wav" and "path/audio.mp3" are paired accordingly. This strict alignment guarantees the script correctly associates files for conversion based on the specified mode.
Adjust Training Parameters: Although the conversion mode operates under the same training principles as generative modelling, you might want to adjust certain parameters in config.py to optimize the conversion process. This could include tuning the PATCH_SIZE and PATCH_LENGTH settings to better accommodate the file sizes commonly encountered in your conversion tasks.
Leverage Pre-trained Weights (Optional): Same as regular generative modelling, if you wish to fine-tune a pre-trained bGPT model, set PRETRAINED_PATH to the location of the pre-trained weights and ensure LOAD_FROM_PRETRAINED=True. To train a model from scratch, simply set LOAD_FROM_PRETRAINED=False.
Start Training for Conversion: When training bGPT in conversion mode, the model learns to map byte sequences from the source format to the target format (or vice versa in bidirectional mode). Execute train-gen.py to start the training process, ensuring that the CONVERSION_MODE is correctly set in your configuration file.

By leveraging the conversion mode, bGPT enables simulating and reverse engineering the behaviors of algorithms or hardware through paired inputs and outputs, opening up new possibilities for data processing and content generation tasks.

Classification

Classification with bGPT leverages the model's ability to understand and differentiate between various types of data at a fundamental level. This involves extracting a global feature from the byte sequence, which is then processed by a classification head. Here's how to approach classification tasks with bGPT:

Prepare Labelled Data: Ensure your dataset consists of labelled data, which can be a mix of different formats. The model distinguishes between data types using the naming convention label.ext, where the label is derived from the filename, specifically filename.split('_')[0]. This means that the label for classification should be clearly reflected in the file name, such as "Business_1.txt". It is crucial to organize your files accordingly to facilitate accurate classification.
Generative Modelling Before Classification (Strongly Recommended): Before embarking on classification tasks, it is highly recommended to perform generative modelling on the same dataset. Starting with weights trained through generative modelling provides a solid foundation for further fine-tuning in classification tasks. To do this, set PRETRAINED_PATH to your generative model weights and ensure LOAD_FROM_PRETRAINED=True. Directly training a classification model from scratch without this pre-training step has been observed to result in significantly poorer performance. When fine-tuning for classification, ensure that WEIGHTS_PATH and LOGS_PATH are set to different locations to prevent overwriting previous models. Note that the classification model will inherit the bGPT's patch-level decoder and discard the byte-level decoder, so it's essential to keep the model parameters unchanged during this phase.
Start Training for Classification: Run train-cls.py to begin the classification training process. The script will utilize the previously set configurations and apply them to your labelled dataset. The model will learn to classify the input data into the defined categories based on the labels extracted from the filenames.

Inference

Inference with bGPT allows you to apply trained models to new data, performing tasks like data conversion or content generation based on the configurations set in config.py. Here's how to conduct inference using the provided settings:

Set Up the Inference Configuration: First, ensure your config.py includes the configuration for inference as shown above. Adjust the parameters according to the task you want to perform:
- INFERENCE_WEIGHTS_PATH: This should point to the location of your trained model weights that you intend to use for inference. For example, weights-conversion indicates the model trained for converting files from one format to another.
- INPUT_EXT and TARGET_EXT: These parameters define the extensions of the input and target files, respectively. In the given configuration, the model expects input files with the .mid extension and will aim to convert them into files with the .abc extension.
- MODE: Determines the mode of inference. convert mode is used for converting files from one format to another, while generate mode is used for generating new content.
- NUM_SAMPLES, TOP_K, TOP_P, and TEMPERATURE: These parameters control the sampling strategy for generation tasks, influencing the diversity and creativity of the output.
Performing Conversion or Generation: Depending on the MODE you've set, the inference process will either convert input files to a new format or generate new content:
- Conversion: In convert mode, ensure your input files (e.g., .mid) are placed in a designated directory. The model will read these files, apply the conversion process, and output files in the target format (e.g., .abc) in the specified output directory.
- Generation: In generate mode, the model will generate samples from scratch. The number of samples to generate is specified by NUM_SAMPLES. The generated samples will be placed in the output directory.
Executing Inference: To start the inference process, run the script inference.py. Make sure the script is configured to use the settings from config.py for conducting the desired inference task.

CPU States Dataset

The CPU States Dataset is integral to exploring and modelling CPU behaviour, including a simplified yet detailed representation of CPU operations through sequences of machine instructions and subsequent register states. This dataset is crucial for training models like bGPT to predict how the CPU state updates with each instruction, demonstrating the model's proficiency in simulating digital processes within hardware environments.

The generation of the CPU States Dataset is facilitated by the cpu-simulation.py script, which allows for the generation, translation, and evaluation of CPU state sequences. The script operates in various modes to accommodate different stages of dataset preparation and model testing:

Generate Mode: This mode is used to generate new instances of CPU states, simulating a wide range of program instructions to create a diverse dataset. It involves specifying the number of registers, memory size, and the range and quantity of instruction sequences to be generated. Example command:
```
python cpu-simulation.py --mode generate --num_registers 10 --memory_size 1024 --dir_path cpu_states --num_min_program 1 --num_max_program 255 --num_train_instance 2100000 --num_test_instance 21000 --random_seed 0
```
This generates training and testing instances with programs containing 1 to 255 instructions, tailored to the specified memory size and register count.
Translate Mode: It is used to convert binary representations of CPU states into a human-readable format. This conversion is essential for manual inspection, debugging, and understanding the dataset's intricacies. Example command:
```
python cpu-simulation.py --mode translate --states_path cpu_states/cpu.bin
```
Evaluate Mode: To assess the accuracy of the model's predictions, the evaluate mode compares the predicted CPU states against the actual states within the dataset. Example command:
```
python cpu-simulation.py --mode evaluate --dir_path cpu_states
```

The CPU States Dataset contains 2.1 million instances, each featuring a 1KB memory block and sequences of 16-byte CPU register states that include a variety of instruction types, such as data movement, logical, and arithmetic operations. This dataset simulates typical CPU behaviour, providing a rich resource for research and development in digital system simulation.

BibTeX

@misc{wu2024language,
      title={Beyond Language Models: Byte Models are Digital World Simulators}, 
      author={Shangda Wu and Xu Tan and Zili Wang and Rui Wang and Xiaobing Li and Maosong Sun},
      year={2024},
      eprint={2402.19155},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

sander-wood
/

bgpt