The accuracy mode
When 130 KB is not enough: a stored linear model over TF-IDF, trained on 75,000 human-labelled AG News rows. It runs in this page, with no server.
Try it
What is this?
The 130 KB page is the minimum-bit hook: 500 teacher-labelled examples and a Naive Bayes fitted in the browser. This page is the other end of the same trade-off: stored weights, trained on 75,000 human labels, and a 1.8 MB artifact. Both score on the same 400 held-out AG News rows.
The model is a TF-IDF (uni- and bigrams, 75k features) + multinomial logistic regression. Coefficients are int8-quantized; vocabulary and IDF are embedded. The accuracy is computed live in your browser. The test set is not part of the model. Fresh-sample check: a separate 400-row test gives 91.0% (embedded 88.5%).
Why human labels? On the same 5,000 rows, a 4-bit Qwen2.5-3B-Instruct teacher (73.4% accurate, 153 unknowns) gave a 71.3% linear student; human labels gave 84.3%. Label noise is the bottleneck, not the student.
| bytes | artifact | accuracy | note |
|---|---|---|---|
| 130 KB | coreset + NB | 78.3% | the minimum-bit hook |
| — | linear TF-IDF (this page) | — | stored weights, human labels |
| 2.2 MB | BERT-tiny 4-bit | 76.75% | stored weights |
| 30 MB | T5-small 4-bit | F1 0.256 | generation |