llama-factory fine-tuning 3 (Dec 13, 2023)

1 Introduction

In this blog, we will use 3 dataset to fine-tuning our model using llama-factory.

2 dataset preparation

2.1 MedQA dataset (address)

The source of the dataset is designed to examine the doctors' professional capability and thus contains a significant number of questions that require multi-hop logical reasoning.

in this dataset, we select question, answer to build our dataset, you can check the following managament code.

click to view the code

from datasets import load_dataset
import os
import json

# Load the dataset
dataset = load_dataset("bigbio/med_qa", "med_qa_en_4options_source")

# Define the save path
save_path = "../medical/MedQA"  # Change this path to your local directory
os.makedirs(save_path, exist_ok=True)

# Function to save data as JSON with specified columns
def save_as_json(data, filename):
    file_path = os.path.join(save_path, filename)
    
    # Modify the data to include only 'question' and 'answer' columns
    data_to_save = [{
        "instruction": "Assuming you are a doctor, answer questions based on the patient's symptoms.",
        "input": item['question'],
        "output": item['answer']
    } for item in data]
    
    # Write the modified data to a JSON file
    with open(file_path, 'w', encoding='utf-8') as f:
        json.dump(data_to_save, f, ensure_ascii=False, indent=4)

# Save the modified data for train, validation, and test splits
save_as_json(dataset['train'], 'train.json')
save_as_json(dataset['validation'], 'validation.json')
save_as_json(dataset['test'], 'test.json')

dataset format

Then we move the dataset (train.json) into llama-factory/data & change it's name as alpaca_med_qa_en.json, then add the following code to llama-factory/data/dataset_info.json

click to view the code

  "alpaca_med_qa_en": {
    "file_name": "alpaca_med_qa_en.json",
    "file_sha1": ""
  },

2.2 MedCQA dataset(address)

intorduction: same as MedQA

In this dataset, we select question(input), answer+explain(output),for introduction part, we set as "Assuming you are a doctor, answer questions based on the patient's symptoms."

click to view the code

from datasets import load_dataset
import os
import json

# Load the dataset
dataset = load_dataset("medmcqa")

# Define the save path
save_path = "../medical/MedCQA"  # Change this path to your local directory
os.makedirs(save_path, exist_ok=True)

# Function to save data as JSON with specified columns
def save_as_json(data, filename):
    file_path = os.path.join(save_path, filename)
    
    # Modify the data to include only 'question' and 'answer' columns
    data_to_save = [{
        "instruction": "Assuming you are a doctor, answer questions based on the patient's symptoms.",
        "input": item['question'],
        "output": (item['opa'] if item['cop'] == 0 else 
                   item['opb'] if item['cop'] == 1 else 
                   item['opc'] if item['cop'] == 2 else 
                   item['opd'] if item['cop'] == 3 else "No valid answer found") + ": " + (item['exp'] if item['exp'] is not None else "No explanation provided")
    } for item in data]

    # Write the modified data to a JSON file
    with open(file_path, 'w', encoding='utf-8') as f:
        json.dump(data_to_save, f, ensure_ascii=False, indent=4)

# Save the modified data for train, validation, and test splits
save_as_json(dataset['train'], 'train.json')
save_as_json(dataset['validation'], 'validation.json')
save_as_json(dataset['test'], 'test.json')

2.3 PubMedQA dataset (address)

In this dataset, we set question as introduction, context as input, final_decion+long_answer as output.

click to view the code

from datasets import load_dataset
import os
import json

# Load the dataset
dataset = load_dataset("pubmed_qa", 'pqa_artificial')

# Define the save path
save_path = "../medical/PubMedQA"  # Change this path to your local directory
os.makedirs(save_path, exist_ok=True)

# Function to save data as JSON with specified columns
def save_as_json(data, filename):
    file_path = os.path.join(save_path, filename)
    
    # Modify the data to include only 'question' and 'answer' columns
    data_to_save = [{
        "instruction": item['question'],
        "input": str(item['context']),
        "output": item['final_decision']  + ": " + item['long_answer']
    } for item in data]

    # Write the modified data to a JSON file
    with open(file_path, 'w', encoding='utf-8') as f:
        json.dump(data_to_save, f, ensure_ascii=False, indent=4)

# Save the modified data for train, validation, and test splits
save_as_json(dataset['train'], 'train.json')
# save_as_json(dataset['validation'], 'validation.json')
# save_as_json(dataset['test'], 'test.json')

3 fintuning commands

3.1 med_qa

click to view the code

CUDA_VISIBLE_DEVICES=1 python src/train_bash.py \
    --stage sft \
    --model_name_or_path ../llama/models_hf/7B \
    --do_train \
    --dataset alpaca_med_qa_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --output_dir ./FINE/llama2-7b-med_qa_single \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

3.2 med_caq

click to view the code

CUDA_VISIBLE_DEVICES=2 python src/train_bash.py \
    --stage sft \
    --model_name_or_path ../llama/models_hf/7B \
    --do_train \
    --dataset alpaca_med_cqa_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --output_dir ./FINE/llama2-7b-med_cqa_single \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

7B

click to view the code

CUDA_VISIBLE_DEVICES=2 nohup python src/train_bash.py \
    --stage sft \
    --model_name_or_path ../llama/models_hf/13B \
    --do_train \
    --dataset alpaca_med_cqa_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --output_dir ./FINE/llama2-7b-med_cqa_single \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16 \
    >> ./logs/fintuning_13B_medcaq_single.log 2>&1 &

13B

3.3 pubmed_qa

click to view the code

CUDA_VISIBLE_DEVICES=3 python src/train_bash.py \
    --stage sft \
    --model_name_or_path ../llama/models_hf/7B \
    --do_train \
    --dataset alpaca_pubmed_qa_en \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --output_dir ./FINE/llama2-7b-pubmed_qa_single \
    --overwrite_cache \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

3.4 command for med_cqa

mixtral

click to view the code

accelerate launch src/train_bash.py \
    --stage sft \
    --do_train \
    --model_name_or_path mistralai/Mixtral-8x7B-v0.1 \
    --dataset alpaca_med_cqa_en \
    --template default \
    --quantization_bit 4 \
    --lora_target q_proj,v_proj \
    --output_dir ../FINE/mixtral-alpaca_med_cqa_en \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss

mistral