llama-2-7b-dpo-v0.1

Model Details

• SFT based on meta-llama/Llama-2-7b-hf with merged alpaca datasets
• DPO: trained on top of SFT model as LoRa Adapter, with merged hh-rlhf data
• PPO: trained on top of dpo model and reward model, with multi-adapters, with PKU-SafeRLHF data for futher RLHF
• Trained with Deepspeed ZeRO-1 + TRL + QLoRA + Flash-Attntion 2

Model and Training Details

• Finetuned from model: meta-llama/Llama-2-7b-hf

• Dataset:

  • SFT (mixed train):
    • yahma/alpaca-cleaned
    • vicgalle/alpaca-gpt4
  • DPO (mixed train):
    • Anthropic/hh-rlhf
    • Unified-Language-Model-Alignment/Anthropic_HH_Golden
  • PPO:
    • PKU-Alignment/PKU-SafeRLHF-10K
    • PKU-Alignment/PKU-SafeRLHF-30K
    • PKU-Alignment/PKU-SafeRLHF

Training Results

Evaluation

The reward score and toxicity scores are computed and compared with PKU-Alignment/PKU-SafeRLHF-30K data on SFT/DPO/PPO models

Compute Infrastructure

The model is trained using 8 * RTX-3090-24GB/A100-PCIE-40GB

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, trust_remote_code=True,)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True,)

tokenizer.pad_token = tokenizer.eos_token
tokenizer.eos_token = DEFINE_EOS_TOKEN
model.config.eos_token = DEFINE_EOS_TOKEN
model.config.eos_token_id = tokenizer.eos_token_id

def format_prompt(question):
    return f"###Question: {question}\n###Answer: "

instruction = "Your text here"
input = format_prompt(instruction)
inputs = tokenizer(input, return_tensors='pt')
output = model.generate(inputs['input_ids'], max_new_tokens=512, do_sample=False, top_p=1)
output = tokenizer.decode(output[0], skip_special_tokens=True)
print(output)

Model Card Authors

Yiyu (Michael) Ren

Model Card Contact

Email: [email protected]

Framework versions

• PEFT 0.8.2