Rohit Ghumare 69c83c0a51 fix: technical accuracy audit across 15 lesson files ...

Fixes 16 issues found by cross-referencing against Microsoft AI-For-Beginners:
- ReLU attribution: Nair/Hinton popularized, not published (Fukushima 1969)
- Momentum variable renamed m_t to avoid collision with Adam's v_t
- MTEB framing: open-source models now match/exceed commercial
- LoRA alpha=2r is community convention, not paper default
- LoRA intro: per-layer 0.78% changed to general "less than 1%"
- Claude 3.5/4 -> Claude 3.5/3.7 (Claude 4 doesn't exist)
- Gemini Ultra -> Gemini 1.5 Pro
- GPT-2 param count: use 124,438,272 consistently (with weight tying)
- FP8 "one GPU" -> "one 80GB GPU" for clarity
- Llama 3 GPTQ: attributed to community, not Meta
- Random forests: "nearly impossible to overfit" -> "highly resistant"
- GMM gaussian_pdf: fixed normalization for d-dimensional data
- Naive Bayes: corrected claim about word absence in Multinomial NB
- Swish: removed unverified Nike logo name origin
- AlphaFold/PyTorch -> LLaMA (AlphaFold uses JAX)
- Llama 3 tokenizer: SentencePiece -> tiktoken-based BPE
- CoT accuracy: 95% -> 91% for zero-shot, 95% with few-shot
- Function calling: fixed garbled Anthropic/OpenAI schema sentence

2026-04-02 12:00:11 +01:00

..

clustering.py

fix: technical accuracy audit across 15 lesson files

2026-04-02 12:00:11 +01:00