eugene
@eugeneyalt
Care a lot, try hard, have fun. @eugeneyan's Id. 🏂
Note to self: • Work hard • Keep learning • Cherish loved ones • Find people who inspire you • Be kind & egoless • Eat healthy, exercise, sleep well • Read & write • Practice gratitude & meditate • Be present • Enjoy food & nature • Don’t sweat the small stuff • Smile
wow had a crazy bug that resetting my codebooks every step was causing epoch count to stop incrementing
looking forward to a weekend of no meetings, no conducting interviews, no doc reviews, no mediating tension between teams and systems, no conducting office hours, and just refactoring rqvae and semantic id training code with cc and long walks in seattle summer with the fam
I had so much joy this morning that I just had to write this and share my gratitude with the world
what a world we live in! I just took a Jupyter notebook that implements an llm-evaluator, provided it as context to a coding assistant, and then ask it to write an evaluator class with specified inputs, outputs, etc. initially, it was verbose with too many class methods, but with…
evals are all you need
Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains 'We introduce Rubrics as Rewards (RaR), a framework that uses structured, checklist-style rubrics as interpretable reward signals for on-policy training with GRPO. Our best RaR method yields up to a relative…
When my model trains well I sleep well lol (note how bad it was over the weekend and more)
Metrics after a 200 epoch run: 0.82 recall@10 and 0.67 NDCG. That's crazy good given the simple, smol SASRec has to learn to: • Create valid 4-token IDs to represent the product • Predict the next item in a session With a 3-level codebook with 256 codes, that's ~17M possible…
it's alive! i'm astounded you can replace item IDs with sequences of tokens (3 from the codebooks above + 1 to avoid collisions) and train a transformer to output tokens IN THE RIGHT ORDER recommend items. there are at least 256^3 (16.7M) combinations and it got it right.…
look at that codebook distribution 👨🍳💋 (thanks @voxmenthe for the tip!)
wow, not sure what Anthropic did, but Claude Code is working superbly with notebooks now, even with sloppy instructions.
not easy to find other folks that work on rqvaes, recsys, and semantic ids
Runs completed! Here's what I learned • we can reduce reconstruction loss with higher commitment weight, but it increases codebook loss. thus, if what you want is good codebooks and semantic IDs, probably not a good idea. the default of 0.25 works well • EMA worked horribly for…
tools for learning and building • zotero: papers, annotations, citations • obsidian: notes, writing, connecting the dots • claude: thinking partner, improve understanding • claude code: implementation, code review (images from work on semantic IDs and residual-quantized…