license: other
license_name: microsoft-research-license
license_link: https://huggingface.co/microsoft/phi-2/resolve/main/LICENSE
This is the merged model for LoRA https://huggingface.co/Yhyu13/phi-2-sft-dpo-gpt4_en-ep1-lora
This model is a dpo improvement to this base model https://huggingface.co/Yhyu13/phi-2-sft-alpaca_gpt4_en-ep1 who achieve better than text-davinci-003 on AlpcaEval judged by ChatGPT.
AlpacaEval
Quote from this discussion https://huggingface.co/microsoft/phi-2/discussions/38
Since phi2 requires remote code which HF open llm leaderboard would not accept at this moment,
I ran phi2 and my dpo to the AlpcaEval benchmark
https://tatsu-lab.github.io/alpaca_eval/
Here is result evaluated by chatpgpt https://github.com/tatsu-lab/alpaca_eval/pull/183
win_rate standard_error n_total avg_length
gpt4 73.79 1.54 805 1365
claude 70.37 1.60 805 1082
chatgpt 66.09 1.66 805 811
wizardlm-13b 65.16 1.67 805 985
vicuna-13b 64.10 1.69 805 1037
guanaco-65b 62.36 1.71 805 1249
oasst-rlhf-llama-33b 62.05 1.71 805 1079
alpaca-farm-ppo-human 60.25 1.72 805 803
falcon-40b-instruct 56.52 1.74 805 662
phi-2-alpaca-gpt4-dpo(new)55.60 1.75 804 4532
phi-2-alpaca-gpt4(new) 54.23 1.75 804 1138
text_davinci_003 50.00 0.00 805 307
alpaca-7b 45.22 1.74 805 396
phi-2(new) 43.79 1.74 805 924
text_davinci_001 28.07 1.56 805 296
phi-2-alpaca-gpt4-dpo is only slightly better than my previous sft phi-2-alpaca-gpt4, when evaluted by chatgpt, but the dpo tuned model outputs significantly longer result!