Rohit Ghumare 15d7a3df3d feat(phase-05/13): question answering systems ...

Extractive QA (RoBERTa-SQuAD2) + retrieval-augmented (RAG) + generative
(closed-book LLM) as three architectures with different failure modes.
Worked examples with SQuAD-style EM and F1 metrics demonstrated in
code, showing how paraphrase punishes exact match.

Teaches the retrieval-reader split, citation-based evaluation, and
refusal calibration as the three production-grade metrics beyond raw
answer accuracy.

Ship artifact: qa-architect skill that refuses closed-book LLM answers
for regulatory questions and flags missing retrieval-recall baselines.

~75 minutes. Prerequisites lesson 05/11 and 05/10.

2026-04-22 17:34:58 +01:00

..

qa.svg

feat(phase-05/13): question answering systems

2026-04-22 17:34:58 +01:00