la-min/myanmar-gpt-health-faq

Model Details

Model Description

This is a model that can answer health-related questions in Myanmar. The model was trained on a small dataset of 800 health-related question and response data points. WYNN747/Burmese-GPT-v3 is the base model, and QLoRA is used to fine-tune it further. The underlying model was too huge to fine-tune and train on Kaggle's free GPUs. As a result, we employ QLoRA to do efficient fine tuning on a based model.

• Developed by: La Min Ko Ko
• Model type: GPT Causual LLM
• Finetuned from model: WYNN747/Burmese-GPT-v3

---

Uses

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM

config = PeftConfig.from_pretrained("la-min/myanmar-gpt-health-faq")
model = AutoModelForCausalLM.from_pretrained("WYNN747/Burmese-GPT-v3")
model = PeftModel.from_pretrained(model, "la-min/myanmar-gpt-health-faq")

For inferencing, use this simple code.

Before running, install the necessary packages.

%pip install accelerate peft bitsandbytes transformers datasets==2.16.0

PEFT_MODEL = "la-min/myanmar-gpt-health-faq"

config = PeftConfig.from_pretrained(PEFT_MODEL)
model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    return_dict=True,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
tokenizer.pad_token = tokenizer.eos_token

model = PeftModel.from_pretrained(model, PEFT_MODEL)

This is for generation config, can adjust

generation_config = model.generation_config
generation_config.max_new_tokens = 300
generation_config.temperature = 0.9
generation_config.top_p = 0.7
generation_config.num_return_sequences = 1
generation_config.pad_token_id = tokenizer.eos_token_id
generation_config.eos_token_id = tokenizer.eos_token_id

def generate_text(model_path, sequence):
    
    model = model_path
    ids = tokenizer.encode(f'{sequence}', return_tensors='pt')
    final_outputs = model.generate(
        input_ids=ids,
        generation_config=generation_config,
    )
    response = tokenizer.decode(final_outputs[0], skip_special_tokens=True)
    print(response)

sequence = "[Q] အားဆေး ဘာလို့သောက်ကြတာလဲ"
generate_text(model, sequence) 

Training Results & Parameters

This is parameters for trainingArguments.

    per_device_train_batch_size=8,
    gradient_accumulation_steps=4,
    learning_rate=2e-5,
    fp16=True,
    save_total_limit=3,
    logging_steps=10,
    output_dir=".....",
    max_steps=260,
    optim="paged_adamw_8bit",
    lr_scheduler_type="cosine",
    warmup_ratio=0.05,
    report_to="none",

Summary

Due to a lack of training data, the model output cannot produce satisfactory results. More data is needed for training.

Framework versions

• PEFT 0.7.1