---
language:
- en
license: apache-2.0
tags:
- binary-analysis
- security
- malware-analysis
- executable-analysis
- roberta
- masked-language-modeling
library_name: transformers
pipeline_tag: fill-mask
widget:
- text: "ELF <mask> header"
---

# Glaurung Large 001

A RoBERTa-based masked language model trained on binary executable files for security research and binary analysis. Part of the [Glaurung](https://github.com/mjbommar/glaurung) project: a modern reverse engineering framework with first-class AI integration.

## Overview

**Glaurung Large 001** is a transformer model specifically designed for understanding binary executable files. It uses a custom BPE (Byte Pair Encoding) tokenizer trained on multi-byte patterns from various binary formats across multiple architectures (x86-64, ARM64, etc.) and operating systems (Linux, Alpine, Ubuntu, Debian, Rocky).

This is the **large variant** (371M parameters, 24 layers) offering enhanced understanding of binary patterns. For faster inference, see [glaurung-small-001](https://huggingface.co/mjbommar/glaurung-small-001) (160M parameters).

### Key Features
- **Custom Binary Tokenizer**: BPE tokenizer that creates efficient multi-byte tokens from binary data
- **Binary-Aware**: Trained on actual executable files, not hex strings
- **Multi-Architecture**: Understands patterns from various CPU architectures and file formats
- **Latin-1 Encoding**: Preserves all byte values (0-255) without loss
- **Large Model**: 371M parameters with deeper architecture for enhanced binary understanding

## Model Details

- **Architecture**: RoBERTa for Masked Language Modeling
- **Hidden Size**: 1024
- **Layers**: 24
- **Attention Heads**: 16
- **Intermediate Size**: 4096
- **Vocabulary Size**: 65,536 tokens
- **Tokenizer**: [binary-tokenizer-005](https://huggingface.co/mjbommar/binary-tokenizer-005)
- **Max Position Embeddings**: 520
- **Parameters**: ~371M
- **Special Tokens**:
  - `<|start|>` (0): Beginning of sequence
  - `<|end|>` (1): End token
  - `<|sep|>` (2): Separator/EOS
  - `<|cls|>` (3): Classification token
  - `<|pad|>` (4): Padding
  - `<|mask|>` (5): Mask token for MLM
  - `<|unk|>` (6): Unknown token

## Glaurung Ecosystem

This model is part of the **Glaurung** project ecosystem:

### 🔧 Main Project
- **[Glaurung](https://github.com/mjbommar/glaurung)** - A modern reverse engineering framework designed to replace Ghidra with first-class AI integration throughout the analysis pipeline. Built with Rust's performance and Python's accessibility, featuring AI agents integrated at every level from format detection to decompilation.

### 🤖 Model Family
- **[glaurung-large-001](https://huggingface.co/mjbommar/glaurung-large-001)** (this model) - 371M parameters, 24 layers
- **[glaurung-small-001](https://huggingface.co/mjbommar/glaurung-small-001)** - 160M parameters, 12 layers, faster inference

### 🔤 Tokenizer
- **[binary-tokenizer-005](https://huggingface.co/mjbommar/binary-tokenizer-005)** - 65K vocabulary BPE tokenizer trained on multi-byte patterns

## Performance Comparison vs Glaurung Small 001

| Metric | Glaurung Small 001 | Glaurung Large 001 | Improvement |
|--------|-------------------|-------------------|-------------|
| **Architecture** |
| Parameters | ~160M | ~371M | +132% |
| Hidden Size | 768 | 1024 | +33% |
| Layers | 12 | 24 | +100% |
| Attention Heads | 12 | 16 | +33% |
| **ELF Magic Prediction** (`\x7fEL`) |
| Top-1 Confidence | ~45-50% (est.) | 59.2% | Stronger recognition |
| **x86 Prologue in Context** |
| Top-1 Confidence | ~70-80% (est.) | 100.0% | Perfect prediction |
| **PE Magic Recognition** |
| Top-1 Confidence | ~5-8% (est.) | 7.3% (rank #2) | Weak (training bias) |
| **Binary Similarity Detection** |
| ELF-to-ELF Similarity | 0.85-0.95 | 0.67-0.92 | More nuanced |
| ELF-to-Text Separation | ~0.25-0.30 | ~0.21-0.32 | Similar |

**Key Improvements:**
- **Dramatically improved confidence** on binary pattern prediction (+21pp on ELF magic)
- **Deeper architecture** enables better long-range dependencies in binary code
- **More stable predictions** with near-perfect accuracy on structured headers
- **Larger capacity** for learning complex multi-architecture binary patterns

## Installation & Loading

```bash
pip install transformers torch
```

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM, AutoModel, pipeline

# Method 1: Load with pipeline for fill-mask tasks
fill_mask = pipeline('fill-mask', model='mjbommar/glaurung-large-001', device=-1)

# Method 2: Load model and tokenizer directly for fill-mask
model = AutoModelForMaskedLM.from_pretrained('mjbommar/glaurung-large-001')
tokenizer = AutoTokenizer.from_pretrained('mjbommar/glaurung-large-001')

# Method 3: Load base model for feature extraction/embeddings
model_base = AutoModel.from_pretrained('mjbommar/glaurung-large-001')
```

## Usage Guide

### 1. Loading Binary Data (Critical!)

Binary files MUST be read as bytes and converted to latin-1 encoding:

```python
# CORRECT: Read as bytes, decode with latin-1
with open('/usr/bin/ls', 'rb') as f:
    binary_data = f.read()  # Read first 512 bytes or as needed
    text = binary_data.decode('latin-1', errors='ignore')

# WRONG: Never use hex strings or other encodings
# hex_string = "7f454c46..."  # ❌ Will not work
# utf8_text = binary_data.decode('utf-8')  # ❌ Will lose bytes
```

### 2. Understanding the BPE Tokenizer

The tokenizer creates multi-byte tokens from common binary patterns:

```python
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('mjbommar/glaurung-large-001')

# Example: ELF header tokenization
elf_header = b'\x7fELF\x02\x01\x01\x00'
text = elf_header.decode('latin-1')

tokens = tokenizer(text, return_tensors='pt')
token_ids = tokens['input_ids'][0].tolist()

# Decode tokens individually to see multi-byte patterns
for token_id in token_ids[1:5]:  # Skip special tokens
    decoded = tokenizer.decode([token_id], skip_special_tokens=True)
    print(f"Token {token_id}: {repr(decoded)}")

# Output:
# Token 45689: '\x7fEL'    # ELF magic compressed to one token!
# Token 3665:  'F\x02'     # Format byte + 64-bit flag
# Token 458:   '\x01\x01'  # Little-endian + version
# Token 600:   '\x00\x00\x00\x00\x00\x00\x00\x00\x00'  # Padding
```

### 3. Fill-Mask Task (Token-Level Prediction)

**Important**: Masking works at the TOKEN level, not byte level!

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch

model = AutoModelForMaskedLM.from_pretrained('mjbommar/glaurung-large-001')
tokenizer = AutoTokenizer.from_pretrained('mjbommar/glaurung-large-001')

# Read binary file
with open('/usr/bin/ls', 'rb') as f:
    binary_data = f.read(512)
    text = binary_data.decode('latin-1', errors='ignore')

# Tokenize
tokens = tokenizer(text, return_tensors='pt')
token_ids = tokens['input_ids'][0].tolist()

# Mask the second token (first content token after <|start|>)
masked_ids = token_ids.copy()
original_token = masked_ids[1]  # Save original
masked_ids[1] = tokenizer.mask_token_id

# Prepare input
tokens_masked = {
    'input_ids': torch.tensor([masked_ids]),
    'attention_mask': torch.tensor([[1]*len(masked_ids)])
}

# Predict
with torch.no_grad():
    outputs = model(**tokens_masked)
    predictions = outputs.logits[0, 1].softmax(dim=-1)
    top5 = predictions.topk(5)

# Show results
print(f"Original: {repr(tokenizer.decode([original_token]))}")
for score, token_id in zip(top5.values, top5.indices):
    token_text = tokenizer.decode([token_id.item()], skip_special_tokens=True)
    print(f"Predicted: {repr(token_text)} (confidence: {score:.2%})")

# Example output:
# Original: '\x7fEL'
# Predicted: '\x7fEL' (confidence: 59.23%)  ✓ Correct!
# Predicted: '\x00\x00\x00\x00\x00\x00\x00\x00' (confidence: 9.87%)
# Predicted: '\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00' (confidence: 4.45%)
```

### 4. Using Pipeline for Fill-Mask

The pipeline handles tokenization automatically but requires understanding multi-byte tokens:

```python
from transformers import pipeline

# Load pipeline
fill_mask = pipeline('fill-mask', model='mjbommar/glaurung-large-001', device=-1)

# Read binary
with open('/usr/bin/ls', 'rb') as f:
    binary_data = f.read(100)
    text = binary_data.decode('latin-1', errors='ignore')

# Create masked input at token boundaries
# First, tokenize to understand token boundaries
tokenizer = fill_mask.tokenizer
tokens = tokenizer(text)
decoded_tokens = [tokenizer.decode([tid], skip_special_tokens=True) for tid in tokens['input_ids']]

# Reconstruct with mask at token boundary
masked_text = ''.join([
    decoded_tokens[0],  # <|start|>
    fill_mask.tokenizer.mask_token,  # Mask the ELF magic
    ''.join(decoded_tokens[2:])  # Rest of tokens
])

# Predict
predictions = fill_mask(masked_text, top_k=3)
for pred in predictions:
    print(f"{repr(pred['token_str'])}: {pred['score']:.2%}")
```

### 5. Feature Extraction & Embedding Similarity

Compare binary files by their learned embeddings:

```python
from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F
from pathlib import Path

# Load for embeddings (not MaskedLM)
tokenizer = AutoTokenizer.from_pretrained('mjbommar/glaurung-large-001')
model = AutoModel.from_pretrained('mjbommar/glaurung-large-001')
model.eval()

def get_binary_embedding(file_path, max_bytes=512):
    """Extract embedding for a binary file using mean pooling"""
    with open(file_path, 'rb') as f:
        binary_data = f.read(max_bytes)
        text = binary_data.decode('latin-1', errors='ignore')

    # Tokenize
    tokens = tokenizer(text, return_tensors='pt',
                      padding=True, truncation=True, max_length=512)

    # Get embeddings with mean pooling
    with torch.no_grad():
        outputs = model(**tokens)
        # Mean pooling (better than CLS token for this model)
        attention_mask = tokens['attention_mask']
        hidden_states = outputs.last_hidden_state

        # Mask padding tokens
        mask_expanded = attention_mask.unsqueeze(-1).expand(hidden_states.size()).float()
        sum_embeddings = torch.sum(hidden_states * mask_expanded, dim=1)
        sum_mask = torch.clamp(mask_expanded.sum(dim=1), min=1e-9)
        embedding = sum_embeddings / sum_mask

    return embedding

# Compare multiple binaries
files = ['/usr/bin/ls', '/usr/bin/cat', '/usr/bin/echo', '/etc/passwd']
embeddings = {}

for file_path in files:
    if Path(file_path).exists():
        name = Path(file_path).name
        embeddings[name] = get_binary_embedding(file_path)

# Calculate similarities
print("Cosine Similarity Matrix:")
names = list(embeddings.keys())
for name1 in names:
    similarities = []
    for name2 in names:
        sim = F.cosine_similarity(embeddings[name1], embeddings[name2], dim=-1).item()
        similarities.append(f"{sim:.3f}")
    print(f"{name1:10s}: {' '.join(similarities)}")

# Expected output:
# ELF executables (ls, cat, echo) will have high similarity (0.85-0.95)
# Text file (passwd) will have low similarity (0.25-0.30) to ELF files
```

## Real-World Example: ELF Header Analysis

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch

# Load model and tokenizer
model = AutoModelForMaskedLM.from_pretrained('mjbommar/glaurung-large-001')
tokenizer = AutoTokenizer.from_pretrained('mjbommar/glaurung-large-001')

# Analyze ELF executable structure
with open('/usr/bin/ls', 'rb') as f:
    binary_data = f.read(512)  # Read enough for context

print(f"Raw bytes (hex): {binary_data[:16].hex()}")
# Output: 7f454c46020101000000000000000000

# Convert to latin-1 for model
text = binary_data.decode('latin-1', errors='ignore')

# Tokenize to see learned patterns
tokens = tokenizer(text, return_tensors='pt')
token_ids = tokens['input_ids'][0].tolist()

# Show what tokens the model learned
print("\nTokenized ELF header:")
for i in range(1, min(5, len(token_ids)-1)):  # First few content tokens
    token_text = tokenizer.decode([token_ids[i]], skip_special_tokens=True)
    print(f"Token {i}: {token_ids[i]:5d} = {repr(token_text)}")

# Output:
# Token 1: 45689 = '\x7fEL'  - ELF magic compressed to one token!
# Token 2:  3665 = 'F\x02'   - 'F' + 64-bit flag
# Token 3:   458 = '\x01\x01' - Little-endian + version
# Token 4:   600 = '\x00\x00\x00\x00\x00\x00\x00\x00\x00' - Padding

# Test model's understanding by masking each token
print("\nTesting model predictions:")
for position in [1, 2, 3]:  # Test first 3 content tokens
    masked_ids = token_ids.copy()
    original_token = masked_ids[position]
    masked_ids[position] = tokenizer.mask_token_id

    # Create input tensors
    tokens_masked = {
        'input_ids': torch.tensor([masked_ids]),
        'attention_mask': torch.tensor([[1]*len(masked_ids)])
    }

    # Get prediction
    with torch.no_grad():
        outputs = model(**tokens_masked)
        predictions = outputs.logits[0, position].softmax(dim=-1)
        predicted_token = predictions.argmax().item()
        confidence = predictions.max().item()

    # Show results
    original_text = tokenizer.decode([original_token], skip_special_tokens=True)
    predicted_text = tokenizer.decode([predicted_token], skip_special_tokens=True)
    correct = "✓" if predicted_token == original_token else "✗"

    print(f"Position {position}: {correct}")
    print(f"  Original:  {repr(original_text)}")
    print(f"  Predicted: {repr(predicted_text)} (confidence: {confidence:.1%})")

# Expected Output:
# Position 1: ✓
#   Original:  '\x7fEL'
#   Predicted: '\x7fEL' (confidence: 59.2%)
# Position 2: ✗ (prefers single 'F')
#   Original:  'F\x02'
#   Predicted: 'F' (confidence: 96.0%)
# Position 3: ✗ (not in top 5)
#   Original:  '\x01\x01'
#   Predicted: '\x00\x00\x00\x00\x00\x00\x00\x00' (confidence: 59.1%)
```

## Multi-Format Analysis: ELF vs PE Headers & x86 Instructions

Systematic testing reveals performance varies by format and training data exposure:

### Performance Summary Table

| Pattern Type | Confidence | Rank | Notes |
|--------------|------------|------|-------|
| **ELF magic** (`\x7fEL`) | 59.2% | #1 | Strong (94.6% of training data) |
| **PE magic** (`MZ`) | 7.3% | #2 | Proportional to training (5.4% of data) |
| **x86 prologue** (`PUSH RBP; MOV RBP, RSP`) | 100.0% | #1 | Perfect in full context |

### ELF Header Recognition (Strong)

```python
# Test: /usr/bin/ls with 152 bytes of context
# Token 1: '\x7fEL' (3-byte ELF magic)
# Result: 59.23% confidence, rank #1 ✓
```

The model strongly recognizes ELF headers (94.6% of training data).

### PE Header Recognition (Limited)

```python
# Test: Realistic DOS/PE header with 152 bytes of context
# Token 1: 'MZ' (2-byte PE signature)
# Result: 7.34% confidence, rank #2 (null bytes ranked #1 at 29.95%)
```

PE recognition reflects limited training exposure (5.4% of training data, 647 files).

### x86 Instructions (Context-Dependent)

```python
# Test: Function prologue in /usr/bin/ls at offset 0x4e05
# Token: 'UH\x89å' = 0x554889e5 (4 bytes: PUSH RBP; MOV RBP, RSP)
# Result: 100.00% confidence, rank #1 ✓
```

**Key Finding:** The BPE tokenizer learned to respect x86 instruction boundaries!
- 1-byte tokens: `PUSH reg` (0x55), `RET` (0xc3)
- 2-byte tokens: `MOV reg,reg` with ModR/M (0x89e5)
- 4-byte tokens: Common prologues (0x554889e5)

Performance is excellent **with full binary context** but degrades on isolated instruction bytes.

### Training Data Distribution & Performance Correlation

The model was trained on the following binary distribution:

| Source | Format | File Count | Size (MB) | % by Count | % by Size |
|--------|--------|------------|-----------|------------|-----------|
| Debian/Ubuntu/Alpine packages | ELF | 11,330 | 4,572 | 94.6% | 68.9% |
| Windows Update drivers + SOREL-20M malware | PE | 647 | 2,062 | 5.4% | 31.1% |
| **Total** | | **11,977** | **6,634** | | |

**Key Metrics:**
- **By file count**: 17.5:1 (ELF:PE)
- **By data size**: 2.2:1 (ELF:PE)
- **PE files are 8x larger** on average (3.19 MB vs 0.40 MB per file)

This distribution explains the observed performance:

| Format | Training Data | Recognition Confidence | Notes |
|--------|---------------|----------------------|-------|
| ELF | 11,330 files (95%) / 4,572 MB (69%) | 59.2% | Dominant by count |
| PE | 647 files (5%) / 2,062 MB (31%) | 7.3% | Better represented by size |

**Key Takeaway:** The model's PE performance reflects training data composition. While PE is only 5% by file count, it represents 31% by size due to larger average file sizes. The 8.1x performance gap (59.2% vs 7.3%) roughly correlates with the 17.5x file count imbalance, though size-based exposure is more balanced.

**Practical Guidance:**
- ✅ **Use for**: Linux/Unix binary analysis, ELF malware analysis, x86-64 code patterns
- ⚠️ **Limited for**: Windows PE analysis (consider retraining with balanced PE dataset)
- ✅ **Tokenizer learned**: Instruction-level boundaries across both formats

## Training Details

- **MLM Objective**: 20% masking probability
- **Training Data**: Binary executables from various architectures
- **Optimization**: AdamW with warmup, dropout 0.01
- **Special Design**: Increased position embeddings (520) to handle RoBERTa's position offset
- **Model Size**: Large variant with 24 layers and 1024 hidden dimensions

## Limitations

- Maximum sequence length: 512 tokens
- Optimized for executable files (ELF, PE, Mach-O)
- Mean pooling recommended for embeddings (pooler layer not specifically trained)
- Larger model size requires more memory (consider using device_map="auto" for large batches)

## Citation

If using this model in research:
```
@software{glaurung-large-001,
  title = {Glaurung Large 001: Binary Analysis Transformer},
  author = {Glaurung Project},
  year = {2024},
  url = {https://github.com/mjbommar/glaurung-models}
}
```