openai-cookbook VS Hacker News API

It sounds like you are trying to tweak embeddings for similarity search. Rather than fine-tune the model's layers, you may want to try training a linear transformation the existing model's output embedding. Openai has a cookbook on how to do that. You will need some data though - but I think you can try it with ~20 pieces of synthetically generated data.

tutorial from OpenAI https://github.com/openai/openai-cookbook/blob/main/examples/Question_answering_using_embeddings.ipynb

OpenAI Cookbook

OpenAI community repo with lots of examples: https://github.com/openai/openai-cookbook

I recognise the huge amounts of effort involved in this and I applaud the moderators for keeping HN an interesting place to be.

That said, I think it's reasonable for us to have visibility on their manual interventions, and this could be easily surfaced via the Hacker News API (https://github.com/HackerNews/API), if the Story JSON included the values of "contro," "bury," and "gag" fields, which are currently opaque to users of the API

See https://medium.com/hacking-and-gonzo/how-hacker-news-ranking... for more discussion on terminology

Google probably stopped updating BigQuery when they started hosting live Hacker News data in Firebase: https://github.com/HackerNews/API

The live nature of the Firebase data is awesome, but the lack of ability to query is a loss.

Nice work! I did something similar with a personal project a few months ago using an open source llm. Also, not sure if you know, there is an api you can use. https://github.com/HackerNews/API

Hi HN, I made a live feed for viewing all of the new items on Hacker News in almost-realtime. It is a single, modern HTML file that doesn't use polling.

It works by establishing a websocket connection to the Hacker News Firebase database, to receive updates every time the HN server updates Firebase, which is about once every 30 seconds. This is very efficient, putting no load on HN's server and using minimal bandwidth.

To make the feed continuous despite the delay, it waits to display each item until exactly 30 seconds before displaying it. I think this is a fair tradeoff, it gives you a sense of how active HN is. For comparision, there are something like 6-7 thousand tweets every second.

Official HN Firebase API: https://github.com/HackerNews/API

Source code: https://github.com/jerbear2008/hn-live/blob/main/index.html

You could use the API: https://github.com/HackerNews/API

You can now get Hacker News data in real time from the Hacker News API powered by Firebase: https://github.com/HackerNews/API

This is great (real time!), but also kind of a pain (38+ millions individual http requests to get the whole thing).

Thankfully there's no authentication or apparent rate limiting. I fumbled my way through downloading the whole thing with curl. I screwed up a few times so made over 70 million requests in total.

Toy analysis of the data I downloaded here: https://public.tableau.com/app/profile/isna/viz/HackerNewsDa...

Get your submissions from this API

https://hacker-news.firebaseio.com/v0/user/ohjeez.json?print...

and then scan the "submitted" articles as described here

https://github.com/HackerNews/API

I have a crawler that sucks down all the posts from HN and then I read it into Pandas and write all sorts of queries. The boggle I have now is that I want to use the same system to (1) make sure YOShInOn never submits duplicate articles, and (2) have accurate vote and comment scores. (1) requires picking up articles as soon as possible, (2) requires waiting two weeks or so until the scores have settled down to what they are going to be. I guess I gotta come back and rescan things in 2 weeks so I have the right scores.

From looking at the Hacker News API, you might be able to use that instead of scraping from the DOM: https://github.com/HackerNews/API#users

You just need to make 38 million requests to the Firebase API document here https://github.com/HackerNews/API. :-)

I actually made many more than that because I screwed up a few times. Might clean up the compiled data and put it on Kaggle (GitHub? Torrent?), or anywhere else if people have suggestions.