204 lines
4.1 KiB
Markdown
204 lines
4.1 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
|
|
license: apache-2.0
|
|
|
|
base_model: HuggingFaceTB/SmolLM2-360M
|
|
|
|
pipeline_tag: text-generation
|
|
|
|
tags:
|
|
- cybersecurity
|
|
- instruction-tuning
|
|
- smollm2
|
|
- causal-lm
|
|
- transformers
|
|
- pytorch
|
|
- llm
|
|
- ai-security
|
|
|
|
datasets:
|
|
- Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset
|
|
|
|
library_name: transformers
|
|
---
|
|
|
|
# smol-lm2-360-instruct
|
|
|
|
smol-lm2-360-instruct is an instruction-tuned language model built by fine-tuning **SmolLM2-360M Base** on a cybersecurity instruction dataset. The objective is to improve the model's ability to answer cybersecurity-related questions while following instruction-style prompts.
|
|
|
|
> **Note:** This is an educational project for learning Instruction Fine-Tuning (IFT) and should not be considered a production-grade security assistant.
|
|
|
|
---
|
|
|
|
# Model Details
|
|
|
|
| Attribute | Value |
|
|
|-----------|-------|
|
|
| Base Model | HuggingFaceTB/SmolLM2-360M |
|
|
| Model Type | Causal Language Model |
|
|
| Fine-tuning | Full Parameter Instruction Fine-Tuning |
|
|
| Domain | Cybersecurity |
|
|
| Framework | Hugging Face Transformers |
|
|
| Training Library | Transformers Trainer |
|
|
| Language | English |
|
|
|
|
---
|
|
|
|
# Training Objective
|
|
|
|
The model was trained using a supervised instruction-tuning objective.
|
|
|
|
Each training example follows the structure:
|
|
|
|
```text
|
|
<|system|>
|
|
You are a cybersecurity assistant.
|
|
|
|
<|user|>
|
|
Explain SQL Injection.
|
|
|
|
<|assistant|>
|
|
SQL Injection is...
|
|
```
|
|
|
|
During training:
|
|
|
|
- The **System** prompt provides behavioral context.
|
|
- The **User** prompt provides the instruction.
|
|
- Loss is computed **only on the Assistant response**, while the System and User tokens are masked from the loss.
|
|
|
|
This enables the model to learn instruction following while conditioning on the provided context.
|
|
|
|
---
|
|
|
|
# Training Configuration
|
|
|
|
| Parameter | Value |
|
|
|-----------|-------|
|
|
| Epochs | 4 |
|
|
| Learning Rate | 2e-4 |
|
|
| Scheduler | Cosine |
|
|
| Warmup | 10% |
|
|
| Batch Size | 2 |
|
|
| Gradient Accumulation | 8 |
|
|
| Precision | FP16 |
|
|
| Optimizer | AdamW |
|
|
|
|
---
|
|
|
|
# Dataset
|
|
|
|
The model was trained on a cybersecurity instruction dataset consisting of:
|
|
|
|
- System prompts
|
|
- User instructions
|
|
- Assistant responses
|
|
|
|
The dataset covers topics such as:
|
|
|
|
- Network Security
|
|
- Malware Analysis
|
|
- Threat Intelligence
|
|
- Secure Coding
|
|
- Incident Response
|
|
- Vulnerability Management
|
|
- Security Best Practices
|
|
- MITRE ATT&CK
|
|
- NIST Cybersecurity Framework
|
|
|
|
---
|
|
|
|
# Example
|
|
|
|
## Prompt
|
|
|
|
```text
|
|
<|system|>
|
|
You are a cybersecurity expert.
|
|
|
|
<|user|>
|
|
Explain the difference between symmetric and asymmetric encryption.
|
|
|
|
<|assistant|>
|
|
```
|
|
|
|
## Output
|
|
|
|
```text
|
|
Symmetric encryption uses a single shared key for both encryption and decryption, making it computationally efficient and suitable for encrypting large amounts of data. Asymmetric encryption uses a public/private key pair, enabling secure key exchange and digital signatures but at a higher computational cost.
|
|
```
|
|
|
|
---
|
|
|
|
# Usage
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
|
model_name = "Aravindan/smol-lm2-360-instruct"
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
model = AutoModelForCausalLM.from_pretrained(model_name)
|
|
|
|
prompt = """
|
|
<|system|>
|
|
You are a cybersecurity assistant.
|
|
|
|
<|user|>
|
|
Explain SQL Injection.
|
|
|
|
<|assistant|>
|
|
"""
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt")
|
|
|
|
outputs = model.generate(
|
|
**inputs,
|
|
max_new_tokens=200,
|
|
temperature=0.7
|
|
)
|
|
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|
```
|
|
|
|
---
|
|
|
|
# Limitations
|
|
|
|
This model:
|
|
|
|
- is intended for educational and research purposes
|
|
- may generate incorrect or outdated cybersecurity advice
|
|
- has not undergone safety alignment comparable to production LLMs
|
|
- should not be used as the sole source for security decisions
|
|
|
|
Always verify security recommendations with authoritative sources.
|
|
|
|
---
|
|
|
|
# Future Improvements
|
|
|
|
- Larger instruction datasets
|
|
- Multi-turn conversation training
|
|
- Preference Optimization (DPO/ORPO)
|
|
- Retrieval-Augmented Generation (RAG)
|
|
- LoRA and QLoRA fine-tuning
|
|
- Evaluation on cybersecurity benchmarks
|
|
|
|
---
|
|
|
|
# Acknowledgements
|
|
|
|
- Hugging Face
|
|
- SmolLM2 Team
|
|
- Transformers
|
|
- Datasets
|
|
- PyTorch
|
|
|
|
---
|
|
|
|
# Citation
|
|
|
|
If you use this model in your work, please cite this repository. |