One Answer, Four Different Scores: Learning LLM Evaluation the Hard WayHands-on LLM evaluation: EM, F1, BLEU, ROUGE, cosine similarity, plus a real online metrics dashboard built with FastAPI, MySQL, and Streamlit.8 min read·shaka-ai.hashnode.devSep 25
Satu Jawaban, Empat Angka Berbeda: Belajar Evaluasi LLM dari NolBelajar evaluasi LLM dari nol: EM, F1, BLEU, ROUGE, cosine similarity, dan dashboard online metrics pakai FastAPI + MySQL + Streamlit.7 min read·shaka-ai.hashnode.devSep 25
Serving Your Own Model with llama.cpp: GPU Offload, Caching, and the Trap of a Stuck Wrong AnswerServing a GGUF model with llama.cpp, caching answers in SQL, and a demo of why a cached wrong answer doesn't just go away on its own.7 min read·shaka-ai.hashnode.devSep 25
Nge-serve Model Sendiri dengan llama.cpp: GPU Offload, Cache, dan Jebakan Jawaban yang NyangkutMenyajikan model GGUF pakai llama.cpp, cache jawaban di SQL, dan demo kenapa jawaban salah yang tersimpan tidak bisa hilang begitu saja.6 min read·shaka-ai.hashnode.devSep 25
What Does Fine-Tuning Actually Learn? Testing QLoRA with Deliberately Controlled DataA controlled QLoRA experiment: format is learned first, facts come later, and brand drift turns out to just be about epoch count.7 min read·shaka-ai.hashnode.devSep 25
