Honest write-ups on fine-tuning, evaluation, computer vision and LLM systems: what I measured, what I got wrong, and what I would do differently.
I fine-tuned a 7B model to 270/271 on a held-out set. Then I trained TF-IDF + logistic regression on the same data and got 269/271. Here is what I changed in how I report results.
How a microscopy video becomes a motility report: detection, tracking, a queue, and where the time actually goes.
Your service's own error metric is probably measured on training data. Here is a rolling-origin backtest recipe, with a naive baseline in the room.
When thousands of users ask almost the same question, don't pay for thousands of generations. A two-level template cache pattern.
Fine-tuning on templated data makes leakage easy to create and easy to miss. The checks I run now, and how I report results when the leakage cannot be removed.
Sentinel drafts equity research with page-level citations, then audits itself for SEBI rules. Here is how I tested it, what I measured, and the number I stopped quoting.