Model Card for t5-small-long-qa

Model Details

Model Description

This model was trained to generate answers for questions out of a given context.

• Developed by: philipp-zettl
• Model type: Transformer (T5)
• Language(s) (NLP): English
• License: M.I.T
• Finetuned from model [optional]: google/flan-t5-small

Model Sources [optional]

Fine-tune of the amazing google/flan-t5-small

Uses

It's intended to use the model to answers for questions from given context. The context should not exceed the model's context length.

Bias, Risks, and Limitations

No bias evaluation was performed on this model.

How to Get Started with the Model

Use the code below to get started with the model.

context = "This is a long text based of multiple concatenated paragraphs."
question = "My question about something mentioned inside the context."

model_inputs = tokenizer([f"question: {question} context: {context}"], max_length=512, padding=True, truncation=True)
input_ids = torch.tensor(model_inputs['input_ids']).to(device)
attention_mask = torch.tensor(model_inputs['attention_mask']).to(device)
with torch.no_grad():
    sample_output = model.generate(input_ids[:1], max_length=85)
    sample_output_text = tokenizer.decode(sample_output[0], skip_special_tokens=True)
    input_text = tokenizer.decode(input_ids[0], skip_special_tokens=True)
    print(f"Sample Input:\n \"{input_text}\"\n\n")
    print(f"Model Output: \"{sample_output_text}\"")

Training Details

Training Data

This model was trained on philipp-zettl/long-qa.

A synthetic data set created from philipp-zettl/qg-tidyqa_squad2 using microsoft/Phi-3-mini-128k-instruct.

The data set was created by prompting Phi-3 using the prompt template

msg = f"""
Answer the following question using the content provided in the context.
Do not answer questions where the answer isn't inside the context.


Question: {sample['question']}
Context: {sample['context']}
"""

After generating synthetic answers, the data set was manually corrected and validated to ensure high quality as well as consistent longer answers than the original data sets.

Training Procedure

Below you can find the full training pipeline used to achieve this fine-tune.

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Base model (e.g., T5-large)
# https://huggingface.co/collections/google/flan-t5-release-65005c39e3201fff885e22fb
model_name = 'google/flan-t5-small'
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Move only the student model to GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

Load dataset

from datasets import load_dataset

# Load dataset
ds = load_dataset('philipp-zettl/long-qa')

# Split the dataset into training and validation
train_dataset = ds['train']
validation_dataset = ds['test']

Preprocessing: tokenize inputs and labels for faster training cycles, i.e. no need for tokenization during training anymore

def preprocess_batch(batch, tokenizer, max_input_length=512, max_output_length=128):
    questions = batch['question']
    contexts = batch['context']
    answers = batch['answer']

    inputs = [f"question: {q} context: {c}" for q, c in zip(questions, contexts)]
    model_inputs = tokenizer(inputs, max_length=max_input_length, padding=True, truncation=True)

    labels = tokenizer(answers, max_length=max_output_length, padding=True, truncation=True)
    model_inputs['labels'] = labels['input_ids']

    return model_inputs

# Tokenize the dataset
train_dataset = train_dataset.map(lambda batch: preprocess_batch(batch, tokenizer), batched=True)
validation_dataset = validation_dataset.map(lambda batch: preprocess_batch(batch, tokenizer), batched=True)

# Set format for PyTorch
train_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])
validation_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])

The train loop

from tqdm import tqdm
from transformers import AdamW, DataCollatorForSeq2Seq
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

torch.cuda.empty_cache()

model_name = 'google/flan-t5-small'
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Move only the student model to GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

# Training parameters
epochs = 50
learning_rate = 3e-5
temperature = 2.0
batch_size = 8
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)

# Create a data collator for padding and batching
data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model)

# Create DataLoaders with the data collator
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, collate_fn=data_collator)
validation_dataloader = DataLoader(validation_dataset, batch_size=batch_size, collate_fn=data_collator)

writer = SummaryWriter(comment='t5-small-long-qa')

# Store losses and learning rates
train_losses = []
val_losses = []
learning_rates = []

print("Starting training...")

# Training loop
for epoch in range(epochs):
    model.train()
    total_loss = 0
    print(f"Epoch {epoch+1}/{epochs}")

    progress_bar = tqdm(train_dataloader, desc="Training", leave=False)

    for step, batch in enumerate(progress_bar):
        # Move student inputs to GPU
        input_ids = batch['input_ids'].to(device)
        attention_mask = batch['attention_mask'].to(device)
        labels = batch['labels'].to(device)

        # Teacher forward pass on CPU
        outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)
        logits = outputs.logits

        # Calculate losses
        loss = outputs.loss  # Cross-entropy loss
        writer.add_scalar("Loss/train", loss, epoch * len(train_dataloader) + step)

        # Backpropagation
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        total_loss += loss.item()

        # Verbose logging
        if step % len(train_dataloader)//10 == 1 or step == len(train_dataloader) - 1:
            progress_bar.set_postfix({
                'step': step,
                'loss': loss.item(),
            })

            # Generate a sample output from the student model
            model.eval()
            with torch.no_grad():
                sample_output = model.generate(input_ids[:1], max_length=50)
                sample_output_text = tokenizer.decode(sample_output[0], skip_special_tokens=True)
                input_text = tokenizer.decode(input_ids[0], skip_special_tokens=True)
                writer.add_text(f"Sample Input", input_text, step)
                writer.add_text(f"Sample Output", sample_output_text, step)
            model.train()

    
    avg_train_loss = total_loss / len(train_dataloader)
    train_losses.append(avg_train_loss)
    learning_rates.append(optimizer.param_groups[0]['lr'])

    # Validation step
    model.eval()
    total_val_loss = 0
    with torch.no_grad():
        for batch in validation_dataloader:
            input_ids = batch['input_ids'].to(device)
            attention_mask = batch['attention_mask'].to(device)
            labels = batch['labels'].to(device)

            outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)
            val_loss = outputs.loss
            total_val_loss += val_loss.item()

    avg_val_loss = total_val_loss / len(validation_dataloader)
    val_losses.append(avg_val_loss)

    writer.add_scalar("AVG Loss/train", avg_train_loss, epoch)
    writer.add_scalar("AVG Loss/val", avg_val_loss, epoch)

    print(f"Epoch {epoch+1} completed. Avg Train Loss: {avg_train_loss:.4f}, Avg Val Loss: {avg_val_loss:.4f}")


print("Training complete.")
writer.close()