GPT-4 Summarization with Chain of Density Prompting

GPT-4 Summarization with Chain of Density Prompting

Chain of Density Prompt:

👊 Become a member:

Get a FREE 45+ ChatGPT Prompts PDF here:
📧 Join the newsletter:

🌐 My website:

Paper:

In this video I test the GPT-4 Summarizer prompt "Chain of Density" from the paper From Sparse to Dense. This is a Prompt Engineering technique you can try.

[davidtindell]

Thanks for this Vid. I have tried "CoD" several times both directly pasting into GPT4 and via Python. I have found the key is how well we identify (via ';' delimiters) the 'Informative Entities'. If these "I.E.'s" are not carefully selected or are too long or there are too many then the summaries are "not so good" !

[BrianMosleyUK]

Excellent little technique. Thanks for introducing and showing how it works.

[brunocgoedert]

Fantastic and interesting video. Thanks for sharing with your public.

[watcher]

Interesting idea. Thanks for sharing.

[technolus]

Cool finding.
That seems quite an improvement over baseline performance.

[BirgittaGranstrom]

I will test the prompt. It is superinteresting to think "density" of a summary and writing in general. Maybe questions if you are a desity writer.

[yoagcur]

I just tried it out on an article and it responded as expected. However, I found it got over-dense with some of the entities included without explanation. My preference for the article I chose was iteration 3 but it is great to have a choice so quickly available. I guess some leeway on summary length could also be incorporated

[kenhtinhthuc]

Interesting. Can the approach be applied to other forms of writing to make them "denser"?

[linnepagusta]

Hi All, I am missing something here. do all the summaries get generated from a single prompt yes? no ?

if yes am i correct in saying we do not pass a generated summary to model to generate the next summary(as in map reduce in Langchain) and all summaries are being generated in one go. In this case i am unclear how the model is looking at previous summary and determining that is it missing some entities, and including it in the next summary especially since all generation is happening in one go.

Thanks for response in advance

[r_unknown]

Where Can i fond the document please ?

[NatPeterson]

Interesting. If you're doing this through OpenAI APIs I'd output to YAML instead of JSON like the paper shows, to save on some token costs.

[yoagcur]

Your website is not currently accessible so none of the links work

[MaliRasko]

the article has to be pretty short for expected CoD to fit in the 8k context window.. I find this prompt method to be pretty useless.. in that regard. I will have to try it out with Claude 100k.

[Stockholm_Syndrome]

Spännande. Det kommer mycket om AI från Universiteten nu.

[crobinso]

How do they expect the model to know what "entity-dense" means when they just made up the phrase? And why did they choose a confusing term like "entity" instead of something like "content"?