Youliang/llama3-70b-instruct-lora-derta-100step

lora_Meta-Llama-3-70B_derta

This model is a fine-tuned version of meta-llama/Meta-Llama-3-70B-Instruct on the Evol-Instruct and BeaverTails dataset.

Model description

Please refer to the paper Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training and GitHub DeRTa. The model is continued train 100 steps with DeRTa on LLaMA3-70B-Instruct.

Input format:

[INST] Your Instruction [\INST]

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 1
• seed: 1
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 2.0

The lora config is:

{
  "lora_r": 96,
  "lora_alpha": 16,
  "lora_dropout": 0.05,
  "lora_target_modules": [
    "q_proj",
    "v_proj",
    "k_proj",
    "o_proj",
    "gate_proj",
    "down_proj",
    "up_proj",
    "w1",
    "w2",
    "w3"
  ]
}

Training results

Framework versions

• PEFT 0.10.0
• Transformers 4.40.0
• Pytorch 2.2.0+cu118
• Datasets 2.10.0
• Tokenizers 0.19.1