cd /news/machine-learning/how-do-you-decide-which-model-to-sta… · home › topics › machine-learning › article
[ARTICLE · art-170459] src=discuss.huggingface.co ↗ pub= topic=machine-learning verified=true sentiment=· neutral

How do you decide which model to start with for a first project?

A first machine-learning project classifying personal email as important or junk succeeded by choosing constraints before models: the task had to run on a laptop, cost nothing per email, and expose why each call was made, which led to TF-IDF with a 100-tree Random Forest running entirely locally, while GPT-3.5 was limited to writing summaries of emails actually opened — one API call per email read rather than one per email that exists. The author's general rule is to start with the smallest thing that can do the job on hardware already owned and to pick by what needs to be inspected rather than leaderboard position, noting that scikit-learn on a local machine teaches the train, evaluate and retrain-from-feedback loop faster than wiring up a hosted endpoint for text classification or tabular prediction. The stated regret is judging accuracy by feel for the first weeks instead of building a small held-out evaluation set on day one, which the author says would have caught bad assumptions earlier.

read1 min views4 publishedNov 16, 2026

My first project was exactly this decision, so here is how it went. The task was classifying my own email as important or junk. I did not pick a model first; I picked the constraint: it had to run on my laptop, cost nothing per email, and let me see why it made a call. That landed on TF-IDF with a Random Forest of 100 trees, all local. The LLM (GPT-3.5) only got the job of writing summaries of emails I actually open, so it is one API call per email read instead of one per email that exists.

The general version of that rule: start with the smallest thing that can do the job on hardware you already have, and pick by what you need to inspect rather than leaderboard position. If your task is text classification or tabular prediction, scikit-learn on your own machine teaches you the whole loop (train, evaluate, retrain from feedback) faster than wiring up a hosted endpoint.

What I would do differently: build a small held-out evaluation set on day one. I judged accuracy by feel for the first weeks, and a real number would have caught my bad assumptions earlier.

── more in #machine-learning 4 stories · sorted by recency
── more on @gpt-3.5 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/how-do-you-decide-wh…] indexed:0 read:1min 2026-11-16 · —