search

huggingface / lighteval

LightEval is a lightweight LLM evaluation suite that Hugging Face has been using internally with the recently released LLM data processing library datatrove and LLM training library nanotron.
MIT License
467 stars 54 forks source link

With chat templates, instructions shouldn't be prepended to system prompt #110

Closed Whadup closed 3 months ago

Whadup commented 3 months ago

https://github.com/huggingface/lighteval/blob/df21407d9f714bde9ecfb4dd8283afdc2150eec3/src/lighteval/few_shot_manager.py#L199

If we use a chat_template, have a system prompt and an instruction, the instruction is prepended to the system prompt message. It should be prepended to the first user message instead.

clefourrier commented 3 months ago

Why not. We are adding it with the system prompt so it plays better with the truncation options iirc. We can provide both options btw, but I don't feel like there is "a single right way" of managing model prompts at the moment.

Whadup commented 3 months ago

Sure, there's not one single right way. Adding it in front of the system prompt feels weird to me tho, particularly for longer system prompts. In this fake example, I think the order of inputs is funky

The following are multiple choice questions on fish. You're a helpful assistant who only answers in rhyme. When you do not understand the question, politely ask for clarifications. Always remain respectful.

Idk, my intuition is that chat-models expect the prompt as the last instruction.

clefourrier commented 3 months ago

Oh true, got your point! Yep that's a bug, should be examples[0]["content"] + instruction!

Whadup commented 3 months ago

Yeah, that seems like a good solution. Alternatively, we could prepend to the first message with role="user".

clefourrier commented 3 months ago

We have both options actually, depending on whether there is a system prompt or not. As mentioned above, they will play differently with truncation