https://cattolatte.github.io/notes/ Machine learning engineer and open-source developer building NLP systems from first principles. Author of Meridian (from-scratch grounded RAG over biomedical literature), Polaris (encoder-side NLP platform), and Zenith (generative NLP library). en-us Mon, 03 Aug 2026 00:00:00 GMT https://cattolatte.github.io/notes/group-leakage-cross-validation/ https://cattolatte.github.io/notes/group-leakage-cross-validation/ Michi warned users about group leakage while its own benchmark command ignored groups entirely. How entity leakage inflates scores, and why it had to be fixed before the API freeze. Mon, 03 Aug 2026 00:00:00 GMT https://cattolatte.github.io/notes/reranker-below-random/ https://cattolatte.github.io/notes/reranker-below-random/ Meridian's cross-encoder pushed Recall@5 to 0.029 — below random ordering. Diagnosing the collapse, and why fallible pipeline stages must degrade gracefully. Sun, 19 Jul 2026 00:00:00 GMT https://cattolatte.github.io/notes/speculative-decoding/ https://cattolatte.github.io/notes/speculative-decoding/ How Zenith uses a small draft model to cut target-model forward passes by 3×+ while producing output identical to greedy decoding. Fri, 10 Jul 2026 00:00:00 GMT https://cattolatte.github.io/notes/weight-tying/ https://cattolatte.github.io/notes/weight-tying/ Sharing the input embedding and output projection — why it works, when it stops working, and how it interacts with logit scale. Tue, 07 Jul 2026 00:00:00 GMT https://cattolatte.github.io/notes/imdb-four-levers/ https://cattolatte.github.io/notes/imdb-four-levers/ What happened when Polaris pulled four classic NLP levers — a transformer, subwords, GloVe, and MLM pretraining — and every one bounced off ~86%. Sat, 04 Jul 2026 00:00:00 GMT https://cattolatte.github.io/notes/rmsnorm/ https://cattolatte.github.io/notes/rmsnorm/ Why modern LLMs dropped mean-centering from LayerNorm, and what benchmarking the two in Zenith showed. Tue, 30 Jun 2026 00:00:00 GMT https://cattolatte.github.io/notes/rope-explained/ https://cattolatte.github.io/notes/rope-explained/ Rotary position embeddings as rotations of query/key pairs — and the training/inference consistency bug that taught me how they really work. Tue, 16 Jun 2026 00:00:00 GMT https://cattolatte.github.io/notes/understanding-bpe/ https://cattolatte.github.io/notes/understanding-bpe/ What BPE actually does, why every modern LLM uses a variant of it, and what implementing a tokenizer from scratch taught me. Tue, 02 Jun 2026 00:00:00 GMT