JeffyPlayground
GitHub ↗
Connecting

Pretrained Classifiers

70 classifiers ready to use. Click a model to try it.

doom_fireLive Demo
Real-time Doom gameplay classifier (24 game-state features)
3 classes24 featuresuser-provided
inbox_routerLive Demo
Train and classify emails into Work, Family, Promo, Notifications
4 classestextMIT
poker_decisionLive Demo
4 AI players play Texas Hold’em using game-state classifiers
4 classes18 featuresMIT
fly_navigationLive Demo
Biomechanical fruit fly forages for food via chemotaxis
3 classes16 featuresMPL 2.0
multilingualLive Demo
Voice command intents in 51 languages (MASSIVE dataset, multilingual encoder)
60 classes51 languagesCC BY 4.0
medicationLive Demo
OCR medication detection: identify drugs + dosage forms from label scans
50 medications10 formsMIT
sms_spam99.1%
SMS spam detection
2 classestextCC BY 4.0
dbpedia96.0%
Wikipedia article ontology classification
14 classestextCC BY-SA 3.0
imdb94.8%
Movie review sentiment (positive/negative)
2 classestextAcademic / non-commercial
banking7794.3%
Banking customer service intent detection
77 classestextCC BY 4.0
ag_news90.6%
News article topic classification
4 classestextAcademic / non-commercial
sst290.1%
Movie review sentiment (positive/negative)
2 classestextStanford academic license
clinc_oos88.4%
Intent detection with out-of-scope
151 classestextCC BY 3.0
massive_intent88.1%
Amazon MASSIVE voice command intents
60 classestextCC BY 4.0
tweet_eval_offensive81.0%
Offensive language detection
2 classestextTwitter TOS / academic
tweet_eval_emotion78.1%
Tweet emotion detection
4 classestextTwitter TOS / academic
emotion75.5%
Text emotion detection
6 classestextAcademic
tweet_eval_sentiment66.2%
Tweet sentiment analysis
3 classestextTwitter TOS / academic
snli65.6%
Natural language inference
3 classestextCC BY-SA 4.0
← Back to catalog

Labels

Info

Try it

Usage

← Back to catalog
doom_fire
Real-time Doom gameplay classifier. Extracts 24 game-state features and decides: FIRE, TURN LEFT, or TURN RIGHT. Logistic regression, no neural network, no GPU.
3Classes
24Features
<1msLatency
CPURuntime
Doom frame
HOLD
100HP
26Ammo
0Kills
0Enemies
1Episode

Decision Log

    ← Back to catalog
    fly_navigation
    A biomechanical fruit fly (NeuroMechFly) searches for food in a virtual arena. At each decision step, Jeffy reads 16 features — food distance, food angle, velocity, heading, and recent action history — and classifies the next move. The fly spirals toward food using chemotaxis, a navigation strategy common in real insects. When food is reached, a new source spawns elsewhere. The classifier was trained on 1,000 simulated foraging episodes.
    3Classes
    16Features
    <1msClassify
    MuJoCoPhysics
    Fly simulation
    WALK
    —Food dist
    —Food angle
    —Speed
    0Foods
    0Step

    Stats

    Classify latency—
    Foods reached0
    Step0
    Heading—

    Decision Log

      ← Back to catalog
      Multilingual Classifiers
      51 per-language voice command classifiers (60 intents each), trained on Amazon MASSIVE in 22 minutes on a CPU. All share a multilingual sentence encoder (paraphrase-multilingual-MiniLM-L12-v2, 384-dim, 50+ languages) — the same architecture, trained on each language’s native data. Each classifier is ~100KB; the 470MB encoder is shared. Because the encoder maps all languages into the same embedding space, you can even pass text in one language to another language’s classifier and it mostly works.
      51Languages
      60Intents
      ~100KBPer classifier
      22minTotal training

      Try it

      Example phrases

      Same intent, different languages. Click a phrase to classify it.
      ← Back to catalog
      Inbox Classifier Demo
      Train an inbox router from 24 labeled examples, then watch it classify new messages in real time. CPU only, no API keys.
      4Classes
      24Train examples
      ~50msLatency
      CPURuntime

      Training Data

      24 labeled examples
      terminal
      pip install jeffy-classify

      Try it

      Usage

      ← Back to catalog
      Medication Scanner
      Two classifiers work together: one identifies the medication name (50 drugs, including brand name → generic mapping), the other detects dosage form. Designed for OCR text from camera scans of medication labels, bottles, and packaging. Handles misspellings, truncated text, ALL CAPS, and mixed formats.
      50Medications
      10Dosage Forms
      96%OCR accuracy
      ~80msLatency

      Live Scan Pipeline

      Medication Label
      Waiting...
      →
      OCR
      Extracted Text
      —
      →
      Classify
      Result
      —

      Try it

      ← Back to catalog
      Poker AI Demo
      Four AI players play Texas Hold'em. Each decision is a Jeffy classifier running on 18 game-state features. No neural network, no GPU.
      4Players
      4Actions
      18Features
      <1msClassify
      0pot
      Alice 1000
      Bob 1000
      Carol 1000
      Dave 1000
      preflop Hand #1

      Current Hand

      Hand History

      How it works

      Each player is a logistic regression classifier trained on 21K simulated poker decisions. Features include hand strength, pot odds, position, stack ratios, and betting patterns. The classifier outputs: FOLD, CHECK, CALL, or RAISE.