Story: Article summarization

jbukuts commented 2 months ago

This story pertains to the generation of summaries from scraped content.

Things to keep in mind:

Model selection may play an important part
Depending on the size of the article content may need to be split and merged to not overflow the context windows of models

jbukuts commented 1 month ago

Jul 25 Notes

Recently tried various models and parameters to see yielded summarization. Quick recap:

mixtral: Good at understanding instructions. Bad output though.
flan-ul20: Reasonable output but it can't understand instructions so it ignores my prompt for the most part.
granite: Decent but not optimal outputs. Understands prompts but outputs not as great as wanted.
llama2-70b: Very bad output
llama3-403b: Very good at understanding the prompt. And results include important numbers and summary of key points within sentence limit given in prompt.

With the release of Meta's new 403b parameter model decided to test that. Very pleased with ensuing results so going to continue on with that model.

JibChainCEO commented 1 month ago

I love the llama3-403b. Lets stay the course with it.

jbukuts / jibchain-poc