yoruba-diacritics-quantized

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

yoruba-diacritics-quantized

This model is a fine-tuned version of Davlan/mT5_base_yoruba_adr on a version of Niger-Volta-LTI, provided by Bunmie-e on huggingface.

Model description

The fine-tuning was performed using the PEFT-LoRa technique, aiming to improve the model's performance on tasks like diacritization restoration and generation.

Key Features:

• Base model: mT5_base_yoruba_adr pre-trained on Yoruba text
• Fine-tuned dataset: Yoruba diacritics dataset from bumie-e/Yoruba-diacritics-vs-non-diacritics
• Fine-tuning technique: PEFT-LoRa

Potential Applications:

• Diacritization restoration in Yoruba text
• Generation of Yoruba text with correct diacritics
• Natural language processing tasks for Yoruba language

Code for Testing:

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

config = PeftConfig.from_pretrained("Professor/yoruba-diacritics-quantized")
model = AutoModelForSeq2SeqLM.from_pretrained("Davlan/mT5_base_yoruba_adr")
model = PeftModel.from_pretrained(model, "Professor/yoruba-diacritics-quantized")
tokenizer = AutoTokenizer.from_pretrained("Davlan/mT5_base_yoruba_adr")

inputs = tokenizer(
    "Mo ti so fun bobo yen sha, aaro la wa bayi",
    return_tensors="pt",
)

device = "cpu" # use your GPU if you have

model.to(device)

with torch.no_grad():
    inputs = {k: v.to(device) for k, v in inputs.items()}
    outputs = model.generate(input_ids=inputs["input_ids"], max_new_tokens=100)
    print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True))

Intended uses & limitations

More information coming

Training and evaluation data

More information coming

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

coming soon.

Framework versions

• PEFT 0.7.2.dev0
• Transformers 4.38.0.dev0
• Pytorch 2.0.0
• Datasets 2.16.1
• Tokenizers 0.15.0