zafir stojanovski

• about
• senior ml engineer @ loka
• co-founded uxo.ai
• master in cs @ uni tuebingen
• passionate about rl, catastrophic forgetting, fair evals
• papers
• momentum-based weight interpolation of strong zero-shot models for continual learning
  interpolate @ neurips 2022 (best paper award)
• open-source
• open-thought/reasoning-gym
  (codeio, matrix manipulation, shortest path, course schedule)
• eleutherai/lm-evaluation-harness
  (lambada, paloma, legalbench, metric tickers, docs)
• natolambert/rlhf-book
  (bradley terry loss, prompt template, margin loss, typos)
• huggingface/transformers
  (blip dynamic inputs, docs)
• fun projects
• laser hockey - my winning entry in an rl tournament
• word game bench - testing llms on word games
• morty - an lstm-based chatbot i built in 2018
• writing
• answers to chip huyen's ml interview questions
• probabilistic machine learning
• posts
• 2024-08-04 key-value (kv) cache
• 2024-08-03 grouped-query attention (gqa)
• 2024-08-03 rmsnorm
• 2024-08-02 gated linear units (glu)
• 2024-07-15 rotary position embeddings (rope)
• 2024-07-13 adamw
• 2024-07-12 elastic weight consolidation (ewc)

x linkedin github