This is a rigorously fine-tuned LoRA adapter for facebook/nllb-200-distilled-600M, heavily optimized for translating Swahili (SWA) to Kalenjin (KLN).

📊 Evaluation Results (SWA -> KLN)

BLEU Score: 40.24
chrF Score: 62.38

🛠️ Technical Details & Training

Hardware: Trained locally (42 Epochs) on 5060 8GB GPU for 13 hours.
LoRA Config: r=64, alpha=128, targeting ["q_proj", "v_proj", "k_proj", "out_proj", "fc1", "fc2"].
Token Strategy: Kalenjin uses "token hijacking" and routes through the luo_Latn token space to prevent catastrophic forgetting that comes with initializing a raw token.

💻 Usage

python

import torch
from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, NllbTokenizerFast

# Load Base
model_id = "facebook/nllb-200-distilled-600M"
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

# Load Adapter
adapter_id = "mutaician/nllb-swahili-kalenjin-v3"
model = PeftModel.from_pretrained(model, adapter_id)

# Load Tokenizer
tokenizer = NllbTokenizerFast.from_pretrained(adapter_id)
tokenizer.src_lang = "swa_Latn" 

text = "Habari yako?"
inputs = tokenizer(text, return_tensors="pt")

target_lang_id = tokenizer.convert_tokens_to_ids("kln_Latn")

with torch.no_grad():
    generated_tokens = model.generate(
        **inputs,
        forced_bos_token_id=target_lang_id,
        num_beams=5,
        early_stopping=True,
        max_length=256
    )

print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0])