One Answer, Four Different Scores: Learning LLM Evaluation the Hard WayThis week's material was about evaluating LLMs: offline metrics (tested against a fixed dataset) versus online metrics (measured from real user interaction). It sounds straightforward on the slides. OSep 25 · 8 min readshaka-ai.hashnode.dev
Satu Jawaban, Empat Angka Berbeda: Belajar Evaluasi LLM dari NolMinggu ini materinya soal evaluasi LLM: offline metrics (diuji pakai dataset tetap) vs online metrics (diukur dari interaksi user sungguhan). Konsepnya kedengaran sederhana di slide. Begitu dicoba senSep 25 · 7 min readshaka-ai.hashnode.dev
Serving Your Own Model with llama.cpp: GPU Offload, Caching, and the Trap of a Stuck Wrong AnswerThe model is fine-tuned. The next question: how do you actually use it — not inside a training notebook, but served like a real API, with a cache so the same question doesn't get recomputed over and oSep 25 · 7 min readshaka-ai.hashnode.dev
Nge-serve Model Sendiri dengan llama.cpp: GPU Offload, Cache, dan Jebakan Jawaban yang NyangkutModel sudah selesai di-fine-tune. Langkah berikutnya: bagaimana cara benar-benar memakainya — bukan di notebook training, tapi disajikan seperti API sungguhan, dengan cache biar tidak menghitung ulangSep 25 · 6 min readshaka-ai.hashnode.dev
What Does Fine-Tuning Actually Learn? Testing QLoRA with Deliberately Controlled DataFine-tuning is often treated as a way to "embed knowledge" into a model — but what actually happens under the hood can be far more layered than that. This post builds a small, controlled experiment — Sep 25 · 7 min readshaka-ai.hashnode.dev
