Closed IamExperimenting closed 5 months ago
Hi, thanks for your interest!
My question is,
Yes, though note gist tokens are specialized for cases where you'd like to compress multiple prompts, but you do not know what they are ahead of time and would like to arbitrary compress multiple prompts without finetuning. If you instead know exactly the prompt you'd like to compress ahead of time and don't mind fine-tuning, you might consider simply fine-tuning to distill away a single prompt using a method like prefix/prompt-tuning or LoRA.
Again, if you have a single very long prompt, you'll probably get the best performance simply distilling away the prompt directly, rather than using a meta-learning method like gist tokens. Gist tokens might perform ok on this task, but I imagine your ability to compress extremely long documents is an empirical question and depends on having a rich enough dataset of long documents + QA pairs so your model can learn general prompt compression.
Hi @jayelm @eltociear ,
I just read your paper "Learning to Compress Prompts with Gist Tokens" . It is an impressive idea.
I thought this approach would help me to solve my problem.
where I'm working on a use-case Question Answering, and I will be passing the static prompt to the LLM.
My question is,
Basically, I have a pdf file which is about 45 pages, and it remains static. I would like ask questions about the pdf, here instruction, context will remain as it is(static) only question will be changing.
please find the example prompt for your reference
Prompt = """ You're a helpful chatbot. Please read the context between and and answer the question accordingly. If the context does not contain any information about the question, please just say I don't know.
Question : what is the display size of iphone 7? Answer: """ Here, only the question will change (basically, user will be asking the question about the document which is context). Here instruction, context remains the same. Question will come from the user and model has to predict/respond with an answer. is it doable with gist token approach?