Rohit Ghumare 97153bad63 fix(phase-19/63): use last non-pad logit for VQA and align docs with reality ...

CodeRabbit:
- VQA prediction read vqa_logits[:, 0, :], which is the first timestep rather
  than next-token-after-the-question. Compute the last non-pad index per
  example and gather the logit there; vqa_em now reflects the model's actual
  answer.
- Prerequisites line pointed at Phase 19 lessons 30-37 (Track B); this lesson
  builds on Track E lessons 58-62. Fixed.
- Docs claimed three eval JSON files were emitted and that evaluate() was
  test-covered; in reality the suite is built in memory and tests target the
  metric helpers and suite shape. Updated docs to describe what ships.

2026-05-26 21:03:18 +01:00

..

main.py

fix(phase-19/63): use last non-pad logit for VQA and align docs with reality

2026-05-26 21:03:18 +01:00

test_main.py

feat(phase-19/63): multimodal-eval

2026-05-26 19:36:33 +01:00