A news classifier that fits in a text file
Hundreds of labeled examples, a few lines of Naive Bayes, and no stored weights. The classifier is fitted in your browser when the page loads. Move the coreset size and watch accuracy against bytes.
Coreset size
training…
Try it
What is this?
The model is the coreset: a few hundred teacher-labeled news articles. The training procedure is a Complement Naive Bayes over binary unigram features. Nothing is pre-trained and no weights are stored — the artifact is the data, and the fit happens here.
This page is not a quantized LLM: there is no 4-bit base here. The Qwen adapter experiments use a 4-bit base; this is the “data as artifact” rung of the ladder.
Accuracy is computed live on 400 held-out AG News rows embedded in this page. The test set is not part of the model. The teacher that labeled the training rows was a 1.5B model; the bits-per-parameter number below counts the coreset against that teacher, not against this page.
| bytes | artifact | accuracy | note |
|---|---|---|---|
| 93 B | date program | 100% exact | different task |
| 341 B | AG News prompt | 73.75% | needs a shared 1.5B base |
| — | this coreset | — | trains here, no base |
| 340 KB | TF-IDF linear | 74.25% | stored weights |
| 2.2 MB | BERT-tiny 4-bit | 76.75% | stored weights |
| 30 MB | T5-small 4-bit | F1 0.256 | generation |
At 2,000 examples the same classifier reaches 82.8% — above the 2.2 MB BERT-tiny (76.75%) on this test set. The point is representation choice: the data can be the smaller artifact, and the shared training procedure does the rest.
Fresh-sample check: a separate 400-row test gives 74.3% at 500 examples and 82.8% at 2,000. The 500-example gap is within the ~4-point noise of a 400-row test; the 2,000-example result matches.
Need more accuracy? The 1.7 MB accuracy mode uses 75,000 human labels and reaches 88.5% on the same 400 rows.
Want your own data? Tiny AI Maker trains a classifier from a CSV in your browser and exports a single HTML file.
One more extreme: AI in a QR code — a 2.2 KB URL with an SMS spam classifier and no weights file. The measured ladder is in the Minimum-Bit Table; the adapter autopsy shows where 4-bit hurts, and the fingerprint survives it. The same SMS model can live in a 10.5 KB spreadsheet, a 1.5 KB favicon, a 1.6 KB WAV, a 4.9 KB SVG or a 3.7 KB PDF model card. The survival matrix shows which of them survives a re-save, re-encode or transcode. The 130 KB coreset itself is also a 175 KB PDF: its examples are in `/TinyAIData` and the page trains them in the browser.