Over the past month I have been working on Open Recommender, an open source YouTube video recommendation system which takes your Twitter feed as input and recommends YouTube-shorts style clips tailored to your interests. I made a video about it if you want a more in-depth introduction.
In this article I want to quickly share my plan for the next month to polish Open Recommender to a beta state. If you prefer, there is a 3 minute breakdown of the roadmap, otherwise feel free to skim the rest of the article.
Tasks
Get Users
Open Recommender Alpha will be me DMing people on Twitter, running the pipeline on their data, sending them recommendations and asking for feedback. The GPT-4 request logs will get auto saved into OpenPipe for fine tuning.
Fine Tune
I will fine tune using a dataset collected from the Open Recommender Alpha. I will also use data collected from running the pipeline over slices of my own twitter data. I will use OpenPipe to do the fine tuning.
I won't bother using user feedback signals / manual dataset filtering or augmentation at this stage, just raw GPT-4 ⇒ Mistral / Llama 7B. The point is to just bring the cost down.
Once the fine tune is done we can test the fine tuned model performance against GPT-4 to check for performance degradation,
Build UI
I will implement a basic YouTube shorts style UI supporting both mobile and web. I don't think I need to bother with auth yet. Open Recommender only uses public data at the moment, so a user's recommendations can just be a public URL that gets DM'd/emailed to them
Lower Priority
Maybe I'll get round to these.
filterShitpostsprompt step to filter out tweets which aren't relevant for making recommendationsMake a PR to OpenPipe to support adding user feedback to the request logs to make dataset filtering easier.
Next Steps
DM me on Twitter (@experilearning) if you want to try the current version of Open Recommender :)
Top comments (0)