|
--- |
|
language: en |
|
tags: |
|
- causal-lm |
|
- secure-model |
|
- merged-lora |
|
license: apache-2.0 |
|
--- |
|
|
|
# Llama-3.1-8b-Instruct-Secure |
|
|
|
This model is fine-tuned with LoRA adapters for secure behavior and low ASR (Attack Success Rate). |
|
|
|
## Model Details |
|
|
|
- **Base Model**: Llama-3.1-8b-Instruct |
|
- **Fine-tuning**: LoRA method |
|
- **Purpose**: Secure language model with defenses against jailbreaking. |
|
|
|
## Training Details |
|
|
|
- **Dataset**: Custom synthetic data |
|
- **Framework**: PyTorch |
|
- **Sharding**: Model is saved in shards of 100MB to ensure compatibility. |
|
|
|
## Usage |
|
|
|
Load the model and tokenizer as follows: |
|
|
|
```python |
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
model_name = "SanjanaCodes/Llama-3.1-8b-Instruct-Secure" |
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
model = AutoModelForCausalLM.from_pretrained(model_name) |
|
|
|
inputs = tokenizer("Your input prompt here", return_tensors="pt") |
|
outputs = model.generate(**inputs) |
|
print(tokenizer.decode(outputs[0])) |
|
|
