Model evaluation: stopping, alignment and ensembles
Use the small KAPT evaluation utilities to examine why the best epoch, sample identity and decision threshold belong in an experimental protocol.
8 min read
Research & engineering notebook
Learning, with a trail.
Questions from real code, small examples to reason with, and decisions worth revisiting.
Notes & reflections
Compiled October 2026Why one order needs three consistency rules
Use a Silk Road teaching order to distinguish request retries, stock contention and state transitions, then turn those distinctions into testable decisions.
8 min read
Interface reflection: search, navigation and accessibility
Use bilingual search and video tabs to examine the relationship between state, focus and reduced motion.
8 min read
From cleaning to annotation: account for every sample
Use the data preparation demo to examine IDs, duplicate conflicts and quarantine reasons, then derive boundaries for annotation records and model inputs.
8 min read
Prompt learning starts with tensor contracts
Trace task prompts, node features and gradient paths; distinguish frozen parameters from disabled gradients, and define what component checks can establish.
8 min read
From permission boundaries to reviewable delivery
Connect resource authorization, workflow states and release evidence through the archive and counselor-system cases.
8 min read
No matching notes. Try another keyword or view all topics.
How to read these notes
These learning summaries were compiled from public code and project documents, with AI assistance in drafting and translation. They distinguish source observations, simplified examples and proposed checks. They are not a dated lab log or evidence of new benchmark results.
More learning repositories and earlier notes