llama3-8b-pku-CITA-Instruct-DPO-Instruct

llama3-8b-pku-CITA-Instruct-DPO-Instruct

Fine-tuned Llama-3.1-8B using CITA (Calibrated Instruction Tuning with Alignment (SFT + DPO + KL regularization)) on the PKU-SafeRLHF dataset for improved safety alignment.

Model Details

• Base Model: meta-llama/Llama-3.1-8B
• Fine-tuning Method: CITA
• Dataset: PKU-SafeRLHF (10,813 samples)
• Training Date: 2025-11-18
• Precision: BF16 (bfloat16)
• Adapter Type: LoRA (r=16, alpha=16, ~168MB)

Training Hyperparameters

• Learning Rate: 5.407819703808436e-06
• Batch Size (per device): 1
• Gradient Accumulation Steps: 8 (effective batch size: 8)
• Warmup Steps: N/A
• Max Steps: N/A
• Weight Decay: 0.010936370520834316
• LR Scheduler: cosine
• Optimizer: adamw_torch
• Beta (DPO temperature): 0.1067277996528356
• Lambda KL (KL penalty): 0.00023457122236804812
• Max Sequence Length: 2048
• Max Prompt Length: 1024

Evaluation Results

• Final Rewards Margins: 7.5190

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "anonymousML123/llama3-8b-pku-CITA-Instruct-DPO-Instruct")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("anonymousML123/llama3-8b-pku-CITA-Instruct-DPO-Instruct")

# Generate
messages = [{"role": "user", "content": "Explain quantum computing"}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(input_ids, max_new_tokens=128, temperature=0.7)

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

Intended Use

• Primary Use: Safety-aligned conversational AI
• Recommended: Instruction following with harm refusal capabilities
• Not Recommended: Medical/legal advice, factual knowledge (use base Llama-3.1 for general tasks)

Limitations

• Fine-tuned on English-only safety dataset (PKU-SafeRLHF)
• May refuse benign requests if phrased similarly to harmful prompts
• LoRA adapter only - requires base Llama-3.1-8B for inference

License

Llama 3.1 Community License (same as base model)

Citation

@misc{llama3_8b_pku_CITA_Instruct_DPO_Instruct_2024,
  author = {User},
  title = {llama3-8b-pku-CITA-Instruct-DPO-Instruct},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/anonymousML123/llama3-8b-pku-CITA-Instruct-DPO-Instruct}}
}

Framework Versions

• Transformers: 4.46.3
• PyTorch: 2.5.1
• TRL: 0.12.1
• PEFT: 0.13.2