README.md

---
language:
- en
license: apache-2.0
task_categories:
- text-classification
tags:
- binary-analysis
- file-type-detection
- mime-type
- magic-bytes
- cybersecurity
size_categories:
- 10K<n<100K
---

# Magic-BERT Binary File Classification Dataset

This dataset contains tokenized binary file samples for training MIME type classification models.
Each sample is a 64KB chunk from the beginning of a file, tokenized using a byte-level BPE tokenizer.

## Dataset Description

- **Total Samples:** 46,450
- **Number of Classes:** 125
- **Sequence Length:** 999999
- **Vocabulary Size:** 32,768

### Splits

| Split | Samples |
|-------|---------|
| Train | 37,111 |
| Validation | 4,591 |
| Test | 4,748 |

## Features

Each sample contains:

| Feature | Type | Description |
|---------|------|-------------|
| `blake2b` | string | Content hash (unique sample ID) |
| `mime_type` | string | MIME type label |
| `source_path` | string | Original file path |
| `source_root` | string | Dataset source |
| `extension` | string | File extension |
| `source_size_bytes` | int64 | Original file size |
| `sample_bytes` | int32 | Sample size (max 64KB) |
| `tokens` | List[int32] | Tokenized input (999999 tokens) |

## MIME Type Distribution (Top 20)

| MIME Type | Count | Percentage |
|-----------|-------|------------|
| application/octet-stream | 2,000 | 4.3% |
| application/vnd.ms-powerpoint | 2,000 | 4.3% |
| audio/flac | 2,000 | 4.3% |
| text/x-c | 2,000 | 4.3% |
| text/plain | 1,999 | 4.3% |
| image/png | 1,964 | 4.2% |
| application/gzip | 1,574 | 3.4% |
| image/svg+xml | 1,486 | 3.2% |
| text/html | 1,454 | 3.1% |
| application/vnd.ms-excel | 1,380 | 3.0% |
| application/javascript | 1,340 | 2.9% |
| application/x-7z-compressed | 1,200 | 2.6% |
| image/gif | 1,156 | 2.5% |
| text/csv | 1,128 | 2.4% |
| image/webp | 1,014 | 2.2% |
| application/zlib | 987 | 2.1% |
| text/x-lisp | 980 | 2.1% |
| application/x-gettext-translation | 965 | 2.1% |
| text/x-c++ | 800 | 1.7% |
| application/zstd | 781 | 1.7% |