You can deploy this model and start earning money today!

mjbommar/
magic-bert-tokenizer-4k

Name: magic-bert-tokenizer-4k
Availability: InStock

Views

No views yet

Magic-BERT Tokenizer (4K vocabulary)

A byte-level BPE tokenizer trained on binary file data for the Magic-BERT project. This tokenizer is designed for binary file classification and analysis tasks.

Features

Vocabulary Size: 4,096 tokens
Type: Byte-level BPE (Byte Pair Encoding)
Training Data: Binary files from diverse sources (executables, documents, archives, media files, etc.)
Encoding: Latin-1 (ISO-8859-1) - treats each byte as a character

Special Tokens

Token	ID	Purpose
`<\|start\|>`	0	Beginning of sequence (BOS)
`<\|end\|>`	1	End of sequence (EOS)
`<\|pad\|>`	2	Padding
`<\|unk\|>`	3	Unknown token
`<\|cls\|>`	4	Classification token
`<\|sep\|>`	5	Separator token
`<\|mask\|>`	6	Mask token (for MLM)

Usage

python

from transformers import AutoTokenizer

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("mjbommar/magic-bert-tokenizer-4k")

# Tokenize binary data (read as latin-1)
with open("some_file.bin", "rb") as f:
    binary_data = f.read()
text = binary_data.decode("latin-1")

# Encode
tokens = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
print(tokens)

Training Details

This tokenizer was trained using the HuggingFace tokenizers library with:

Byte-level BPE algorithm
Training on ~100K+ binary file samples
Diverse file types: executables, documents, archives, media, source code, etc.

Model

This tokenizer is designed to be used with Magic-BERT models for binary file MIME type classification. See the main model repository for more details.

License

Apache 2.0