juakazike/multilingual-bias-corrector-v1

JuaKazi Multilingual Bias Corrector v1

Seq2seq gender bias correction model covering 6 languages. Fine-tuned from castorini/afriteva_v2_base on ~10K correction pairs.

Usage

Input format: correct bias {lang}: {biased sentence}

Where lang is one of: sw, ha, zu, ki, fr, en

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("juakazike/multilingual-bias-corrector-v1")
model = AutoModelForSeq2SeqLM.from_pretrained("juakazike/multilingual-bias-corrector-v1")

def correct(text, lang):
    prompt = f"correct bias {lang}: {text}"
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=128)
    out = model.generate(**inputs, max_new_tokens=128, num_beams=4)
    return tokenizer.decode(out[0], skip_special_tokens=True)

correct("The chairman will lead the board meeting.", "en")
# -> "The chair will lead the board meeting."

Validation BLEU (val set, 10% held out per language)

Language	Pairs	BLEU
Swahili (sw)	1,586	17.7
Hausa (ha)	1,917	4.1
Zulu (zu)	1,931	0.6
Gikuyu (ki)	867	4.0
French (fr)	636	30.8
English (en)	3,464	38.6