ahczhg/llama-3.2-1b-aegis-safety-lora

Llama-3.2-1B Fine-tuned on Aegis AI Content Safety

Model Description

This model is a fine-tuned version of Meta's Llama-3.2-1B, specialized for content safety and moderation tasks. Llama-3.2 represents Meta's latest advancement in efficient language models, offering excellent performance in a compact 1B parameter size.

This model was fine-tuned using LoRA (Low-Rank Adaptation) on the NVIDIA Aegis AI Content Safety Dataset 2.0, which contains diverse examples of safe and unsafe content across multiple categories.

Model Details

• —Base Model: meta-llama/Llama-3.2-1B
• —Fine-tuning Method: LoRA (Low-Rank Adaptation)
• —Dataset: NVIDIA Aegis AI Content Safety Dataset 2.0
• —Training Samples: 2,000 carefully selected samples
• —Language: English
• —License: Apache 2.0

Capabilities

• —Advanced content safety analysis
• —Multi-category threat detection
• —Context-aware safety classification
• —Nuanced harmful content identification
• —Safety-first response generation

Intended Use Cases

• —Enterprise content moderation
• —Real-time chat safety monitoring
• —Content policy enforcement
• —Community guidelines compliance
• —Automated content review systems

Training Configuration

LoRA Parameters

• —Rank (r): 16
• —Alpha: 32
• —Dropout: 0.05
• —Target Modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj

Training Hyperparameters

• —Learning Rate: 2e-4
• —Batch Size: 4 (per device)
• —Gradient Accumulation Steps: 4
• —Effective Batch Size: 16
• —Epochs: 3
• —Optimizer: AdamW (8-bit paged)
• —LR Scheduler: Cosine with warmup
• —Warmup Ratio: 0.1
• —FP16 Training: Yes
• —Max Sequence Length: 512

Usage

Installation

pip install transformers torch peft

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "ahczhg/llama-3.2-1b-aegis-safety-lora"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Example: Content safety check
prompt = "### Instruction:\nAnalyze this content for safety: 'Your text here'\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        temperature=0.7,
        do_sample=True,
        top_p=0.95
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Advanced Usage with Pipeline

from transformers import pipeline

# Create text generation pipeline
generator = pipeline(
    "text-generation",
    model="ahczhg/llama-3.2-1b-aegis-safety-lora",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Generate safety analysis
result = generator(
    "### Instruction:\nIs this content safe? 'Hello, how are you?'\n\n### Response:\n",
    max_new_tokens=128,
    temperature=0.7,
    do_sample=True
)

print(result[0]['generated_text'])

Evaluation

The model was evaluated on a held-out test set from the Aegis AI Content Safety Dataset. Key metrics include:

• —Perplexity on validation set
• —Content safety classification accuracy
• —False positive/negative rates for harmful content detection

Limitations

• —The model is primarily trained on English language content
• —Performance may vary on domain-specific or highly technical content
• —Should be used as part of a comprehensive content moderation system, not as the sole decision-maker
• —May require fine-tuning for specific use cases or content domains
• —The model's outputs should be reviewed by human moderators for critical applications

Ethical Considerations

• —This model is designed to assist in content safety and moderation tasks
• —It should not be used to censor legitimate speech or suppress diverse viewpoints
• —Decisions about content moderation should involve human oversight
• —The model may reflect biases present in the training data
• —Users should implement appropriate safeguards and appeal processes

Training Data

The model was fine-tuned on the NVIDIA Aegis AI Content Safety Dataset 2.0, which includes:

• —Diverse examples of safe and unsafe content
• —Multiple categories of potentially harmful content
• —Balanced representation of safe content
• —Real-world scenarios and edge cases

Citation

If you use this model in your research or applications, please cite:

@misc{llama_3.2_1b_aegis_safety,
  author = {ahczhg},
  title = {Llama-3.2-1B Fine-tuned on Aegis AI Content Safety},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/ahczhg/llama-3.2-1b-aegis-safety-lora}},
  note = {Fine-tuned on NVIDIA Aegis AI Content Safety Dataset 2.0}
}

Acknowledgments

• —Base model by the original authors: meta-llama
• —Dataset provided by NVIDIA
• —Fine-tuning performed using HuggingFace Transformers and PEFT libraries

Contact

For questions, issues, or feedback, please visit the model repository.

Model Card Authors

• —ahczhg

Model Card Contact

• —https://huggingface.co/ahczhg