mirror of
https://github.com/rohitg00/ai-engineering-from-scratch.git
synced 2026-10-02 01:54:39 +08:00
Rohit Ghumare
c992fce627
- L01 grid_shape: guard non-divisible image/patch; SigLIP config 384 -> 378 - L01 zip strict=True (Ruff B905) - L08 skill: threshold text ±30% -> below 70% / above 130% - L09 docs: 1280x720 example -> 1120x672 (divisible by 28) - L09 skill: 1344x1344 = 1,806,336 pixels (was 2,116,800) - L12 docs: reconcile CFG contradiction (training vs inference) - L12 docs: video token math 64*64*8 = 32,768 (was 1024) - L14 docs: bits-to-bytes conversion ~1.75 KiB (was 14 kB) - L23 skill: split latency into retrieval SLA vs full-answer SLA - L25 critical: success condition compares expected_page, not done sentinel