Mwau/ugalibot-micro-topic-classifier

UgaliBot Micro-Topic Classifier

Fine-tuned distilbert/distilbert-base-uncased for micro-topic classification of short object descriptions into 94 fine-grained classes (Kikamba names).

Performance (held-out test set)

• Accuracy: 0.7979
• F1 (weighted): 0.7701

Labels

Abaya, Biogas, Bodaboda, Chai, Chapati, Dera, Fort Jesus, Hijab, Iko, Ikonĩ, Ikwa, Isyo, Ivanda, Juakalĩ, Juisi, Kachumbari, Kamii, Kamongo, Kanisa, Kit Mikayi, Kyondo (basket carried on the head), Kĩkaangi, Kĩlumĩ, Kĩsululu, Kĩsũkũũ, Kĩthembe, Kĩthima (well), Kĩtũngũũ (Onion), Kĩvĩla (chair), Lelũ ( road), Leso, M-Pesa, Maalu (Potato), Maandamano (Protest), Maandazi, Maembe (mango), Makwasĩ (Sweet Potato), Malondu (sheep), Malũlũ (jerry cans, water cans), Mama mboga, Masũa (beaded bracelets), Matatũ (matatu), Mbemba (maize), Mbuta (Nile Perch), Mbĩlĩka (kettle), Mbũi (goat), Mkokoteni (handcart), Mwatũ (beehive), Mũomo wa Omani (omani doors), Mũsikiti (Mosque at Gede Ruins), Mũsuko (braids), Mũĩmi (Farmer), Nathi (coconut), Ndata (walking stick), Ndeengũ (cooked or uncooked greengrams), Ndĩĩ (mortar & pestle), Ngege (Luo) (Tilapia), Ngima (ugali), Ngovia (hats), Ngũ (firewood), Ngũkũ (Chicken), Ngũlũ (jaggery, cane sugar), Ngũlũe (Pigs), Njũkũ (groundnuts), Nyama Choma, Nyanya (Tomato), Nyũmba ya mavati (house made exclusively of iron sheets), Nyũmba ya ndaka (modern earthen house with corrugated iron roofing), Omena. Omena are culturally relavant., Owalo, Sandara (Tamarind), Siatũ (Shoes), Soda, Soko (Market), Sukuma (kales), Taa (lamp), Tuktuk, Vasco da Gama Pillar, Vidaka (arches for displaying valuables at gede ruins), Vundi (mason), Yĩembe (hoe), Yĩiu (banana tree), Yĩiya ya Victoria (Lake Victoria), Yĩiyu ya mũlalu (Ndizi tamu), ikundi (Passion fruit), Ĩnanasi, Ĩsilia, Ĩtalũ, Ĩtikitiki, Ĩvuti, Ũsũũ (uji, porridge), Ũtuti (broom), Ũvanga (machete), Ũvyũ (Khanjar: Oman curved dagger)

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Mwau/ugalibot-micro-topic-classifier")
model = AutoModelForSequenceClassification.from_pretrained("Mwau/ugalibot-micro-topic-classifier")

inputs = tokenizer("your object description here", return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits
pred_id = logits.argmax(-1).item()
print(model.config.id2label[pred_id])