nicobrenner / jeffy

# Jeffy
Pretrained classifiers you can run and retrain on CPU — text, game state, or any numeric features.
**jeffyclassify.com** · **Live Playground** · **GitHub**
## Try it
 [code] 
Or try without installing (requires uv):
 [code] 
Open http://localhost:8400, pick a classifier, and paste one of these:
| Classifier | Try this text |
|------------|---------------|
| banking77 | I was charged twice for the same transaction |
| sms_spam | WINNER! You have been selected for a free cruise. Reply YES to claim. |
| ag_news | The Federal Reserve raised interest rates by 25 basis points on Wednesday |
## Pretrained capabilities
68 classifiers ship with the package — 17 English/general-purpose and 51 multilingual intent classifiers covering languages from Afrikaans to Vietnamese. Weights are logistic regression coefficients (derived model parameters, not copies of training data). Source datasets and licenses are documented in `ATTRIBUTION.md`.
### English / general-purpose
| Task | What it does | Classes | Test Acc | Test F1 |
|------|-------------|---------|----------|---------|
| sms_spam | SMS spam detection | 2 | 99.1% | 98.0% |
| doom_fire | Doom game-state decisions (fire/turn) | 3 | 100.0% | 100.0% |
| fly_navigation | Drosophila connectome navigation | 3 | 98.0% | — |
| dbpedia | Wikipedia article category | 14 | 96.0% | 95.9% |
| imdb | Movie review sentiment (long text) | 2 | 94.8% | 94.8% |
| banking77 | Banking customer intent | 77 | 94.3% | 94.3% |
| ag_news | News topic (world/sports/business/tech) | 4 | 90.5% | 90.5% |
| sst2 | Movie review sentiment | 2 | 90.1% | 90.1% |
| poker_decision | Texas Hold'em poker decisions | 4 | 88.0% | — |
| clinc_oos | Voice assistant intent + out-of-scope | 151 | 88.4% | 92.1% |
| massive_intent | Smart home voice commands (English) | 60 | 88.1% | 86.4% |
| tweet_eval_offensive | Offensive language | 2 | 81.0% | 74.8% |
| tweet_eval_emotion | Tweet emotion | 4 | 78.1% | 74.7% |
| emotion | Text emotion (6 emotions) | 6 | 75.5% | 67.8% |
| inbox_router | Email inbox routing | 4 | 70.8% | 70.8% |
| tweet_eval_sentiment | Tweet sentiment (3-way) | 3 | 66.2% | 65.7% |
| snli | Natural language inference | 3 | 65.6% | 65.2% |
### Multilingual (51 languages)
51 intent classifiers trained on Amazon MASSIVE — 60 voice-command intents per language, using frozen multilingual embeddings + per-language logistic regression. ~100 KB per classifier, no GPU needed.
With LaBSE embeddings (768-dim), mean accuracy is **83.3%** across all 51 languages — only 1.8pp below fine-tuned XLM-R (85.1%). All languages above 70%, 47/51 above 80%.
| Tier | Languages | Accuracy range |
|------|-----------|---------------|
| Tier 1 (≥85%) | fr, en, ru, pl, fa, hu, sv, ja, hi | 85.1–86.5% |
| Tier 2 (80–85%) | pt, id, az, nl, ms, tr, da, es, ko, ro, fi, sl, zh_cn, sq, nb, it, lv, mn, af, de, el, vi, te, ml, he, bn, tl, is, ur, hy, kn, th, ta, am, my, zh_tw, cy, sw | 80.5–85.0% |
| Tier 3 (<80%) | jv, ka, ar, km | 76.4–78.8% |
Our encoder com