Qwen3-4B-Instruct-2507-Sherlock-LoRA

Qwen3-4B-Instruct-2507 Fine-Tuned (LoRA) - Sherlock

A LoRA-fine-tuned version of Qwen3-4B-Instruct-2507 specialized for cybersecurity investigation and infostealer log analysis.

📊 Model Details

📈 Training Configuration

Epochs: 3
Batch Size: 2 (per device) × 4 (gradient accumulation) = 8 (effective)
Learning Rate: 2e-5
Optimizer: AdamW (fused)
Scheduler: Cosine
GPU: NVIDIA H100 80GB
Training Time: ~3.5 hours
Total Steps: 2,361

📉 Performance

Validation loss progression shows steady convergence:

• Step 500: 0.877763
• Step 1000: 0.768739
• Step 1500: 0.696747
• Step 2000: 0.669923 ← Best Model ✓

🗂️ Repository Contents

Model Files

• adapter_model.safetensors - LoRA weights (505 MB)
• adapter_config.json - LoRA configuration
• tokenizer.json - Tokenizer files
• training_args.bin - Training configuration
• chat_template.jinja - Chat template

Training Data

• train.jsonl - 6,294 training examples
• validation.jsonl - 699 validation examples
• data_stats.json - Dataset statistics and analysis

Training Scripts

• train_qwen3.py - Main training script
• chat_with_model.py - Interactive chat interface
• upload_to_hf.py - HuggingFace upload script

Configuration & Documentation

• training_config.yaml - Complete training hyperparameters
• requirements.txt - Python dependencies
• README.md - This file

🚀 Usage

Load and Use the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")

# Load LoRA weights
model = PeftModel.from_pretrained(
    model,
    "covertlabs/Qwen3-4B-Instruct-2507-Sherlock-LoRA"
)

# Generate text
messages = [
    {
        "role": "system",
        "content": "I'm Sherlock, your investigative assistant. I have access to a database of infostealer logs..."
    },
    {
        "role": "user",
        "content": "Investigate this email: [email protected]"
    }
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
outputs = model.generate(**inputs, max_length=2048)
print(tokenizer.decode(outputs[0]))

Interactive Chat

# Download this repository
git clone https://huggingface.co/covertlabs/Qwen3-4B-Instruct-2507-Sherlock-LoRA

cd Qwen3-4B-Instruct-2507-Sherlock-LoRA

# Install dependencies
pip install -r requirements.txt

# Run interactive chat
python chat_with_model.py

Fine-Tune Further

# Use the provided training script to fine-tune on additional data
python train_qwen3.py

📋 Dataset Information

Domain

Cybersecurity investigation assistant for infostealer log analysis

Statistics

• Total Examples: 6,993
• Training Split: 90% (6,294 examples)
• Validation Split: 10% (699 examples)
• Total Messages: 71,424
• Avg Messages per Conversation: 10.21
• Avg Conversation Turns: 4.61
• Estimated Tokens: ~7.66M
• Message Length Range: 4 - 3,866 characters

Data Quality

• ✓ 100% of examples include system prompts
• ✓ 100% follow consistent message structure
• ✓ All examples properly formatted for instruction tuning
• ✓ Thinking blocks removed for Instruct model compatibility
• ✓ Tool call patterns consistent throughout

🔧 Requirements

Python 3.10+
PyTorch 2.1+
transformers 4.40+
peft 0.9+
datasets 2.17+
accelerate 0.27+

Full requirements: See requirements.txt

🎯 Use Cases

This fine-tuned model excels at:

• Analyzing infostealer logs and breach data
• Pattern matching across compromised credentials
• Security investigation and threat analysis
• Cybersecurity incident response
• Fraud detection and investigation
• Multi-step reasoning about security threats

📝 Training Details

The model was trained using:

• Method: Low-Rank Adaptation (LoRA)
• Framework: HuggingFace Transformers
• Hardware: NVIDIA H100 80GB GPU
• Data: 6,294 domain-specific examples
• Optimization: BF16 precision with fused AdamW
• Checkpointing: Saved every 500 steps
• Best Model: checkpoint-2000

💾 Model Size

• LoRA Adapter: 505 MB
• Base Model: ~8 GB (requires separate download)
• Total Repo Size: 520 MB (without base model)

⚠️ Important Notes

1. Base Model Required: You need the base Qwen3-4B-Instruct-2507 model to use this LoRA adapter
2. Merge Option: To merge LoRA weights with base model:

   model = model.merge_and_unload()
   model.save_pretrained("./merged_model")
   

3. Memory Requirements: Minimum 12-16GB VRAM for inference (with 4-bit quantization)
4. Context Length: Model supports up to 262K tokens, but trained examples used 4K context

📄 License

This model is based on Qwen3-4B-Instruct-2507. Please refer to the base model's license for usage terms.

🙏 Credits

• Base Model: Alibaba Cloud Qwen Team
• Training Framework: HuggingFace Transformers
• LoRA Implementation: PEFT library
• Fine-tuning Data: Proprietary cybersecurity dataset

📞 Support

For issues or questions, please refer to:

• HuggingFace Documentation
• PEFT GitHub
• Transformers GitHub

---

Repository Created: December 6, 2025 Last Updated: December 6, 2025 Model Status: ✓ Ready for Production