Phi-2 responses

Question

Hello,

I appreciate having access to the Phi-2 model through huggingface and Azure. So far I have enjoyed working with the model and was hoping I could use it in a research project. However recently I ran into some interesting responses and I am wondering if that's standard behavior. Please see attached:

As you can see the model does answer the question however it appends extra unprompted text at the end. My code so far does not include any form of pipelining or message history.

Is this expected behavior? I understand there is a disclaimer about verbosity and I was wondering if this is what it was talking about.

Answer 1

Hi @Christopher Watson,

Thank you for providing more details about your query. Based on the information you provided, it seems like the Phi-2 model is generating extra text and responses following its first answer to your prompt within a single turn. This behaviour is consistent with the model's disclaimer about verbosity.

Phi-2 being a base model often produces irrelevant or extra text and responses following its first answer to user prompts within a single turn. This is due to its training dataset being primarily textbooks, which results in textbook-like responses.

Please look into the Limitations of Phi-2 or Microsoft Phi-2

Also see this: Phi-2

I hope you understand! Thank you.

---

If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.