TransportGPT

NCS-Winner-Club

As part of NUS-NCS Innovation Hackathon, 2024

Contributors

Anastasia Goh, Alden Sio, Dylan Lo, Li Shuyao, Xu Ziqi, Zhu Yi Cheng

How to Use the Model as an External User

As an external user, leveraging the fine-tuned model for your applications is straightforward. Follow the steps below to integrate and utilize the model effectively:

Step 1: Installing Dependencies

Ensure you have Python and the necessary libraries installed. You will need all the libraries within the requirements.txt file, which can be installed via pip:

pip install requirements.txt

Step 1.5: Installing Fine-Tuned model

Ensure you download the Checkpoint (updated model) into any portion within your drive. Save the file path.

Step 2: Loading the Model

You can load the fine-tuned model directly using the Transformers library. Replace your_model_path with the actual path where the fine-tuned model is hosted:

from transformers import T5ForConditionalGeneration, T5Tokenizer

model_path = "your_model_path" # Replace this with the path to the fine-tuned model
model = T5ForConditionalGeneration.from_pretrained(model_path)
tokenizer = T5Tokenizer.from_pretrained(model_path)

Step 3: Preparing Your Input

Prepare the text you want to analyze or process. Ensure it's in a format compatible with the model's expectations:

text_to_process = "Your input text here"
inputs = tokenizer(text_to_process, return_tensors="pt")

Step 4: Generating Predictions

With the model and inputs ready, you can now generate predictions:

outputs = model.generate(**inputs)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Step 5: Interpreting the Results

The output will be your model's interpretation or response based on its fine-tuning. Analyze the results as needed for your application.

Fine-Tuning Guide for Emergency Incident Model

This is our guide on how we fine-tuned the "google/flan-t5-base" model for emergency incident reporting. Below is a generic sequence of events that outlines our fine-tuning process:

1. Installation and Importing Libraries

Firstly, we begin by installing and importing necessary libraries and models. For this project, we utilized "google/flan-t5-base" from HuggingFace.

2. Instantiating the Model

We then instantiate the base Google FLAN model for further processing.

3. Dataset Loading and Preprocessing

The dataset is loaded and preprocessed through tokenization. We specifically allow contextual words like "no", "don't", etc., to handle prompts such as "no one is injured" or "don't need to send ambulance".

4. Tokenization into Dictionary Format

Our dataset is further tokenized into a dictionary format, which is a requirement for this model. For our case, keys such as 'input_ids', 'attention_mask', 'labels' are essential for training.

5. System Prompt and Labeling

We add a system prompt, "extract structured details:", and attach labels to the respective columns. This data is then split into training and testing samples.

6. Converting Texts into Embeddings

Text data is converted into embeddings to be processed by the model.

7. Global Training Parameters

Next, we decide on global parameters for training, which mostly depend on computational power. Here are some key parameters:

• L_RATE (Learning Rate): Determines the adjustment rate of network weights with respect to the loss gradient. A smaller value indicates slower adjustments.
• BATCH_SIZE: Specifies the number of samples processed before updating the model's internal parameters.
• PER_DEVICE_EVAL_BATCH: Defines the number of samples processed at once during model evaluation. Usually equal to BATCH_SIZE.
• WEIGHT_DECAY: A regularization technique to prevent overfitting by penalizing larger weights.
• SAVE_TOTAL_LIMIT: Specifies the maximum number of model checkpoints to save.
• NUM_EPOCHS: The number of times the entire dataset passes through the model.

8. Training

With the parameters set, we proceed to train the model using .train() method.

9. Model Checkpointing

After training, we obtain the desired checkpoint (the one with the least loss) and store it. This model can then be loaded using:

last_checkpoint = "./results/checkpoint-500"
finetuned_model = T5ForConditionalGeneration.from_pretrained(last_checkpoint)
tokenizer = T5Tokenizer.from_pretrained(last_checkpoint)

10. Testing the Model

Finally, we test the fine-tuned model with prompts to evaluate its performance. For example:

incident_report = "Hello police, there is an accident near me at Information Technology NUS, Street 2. A bus collided with a Taxi, 3 people are severely injured, there is a fire. Students are calling for help, Lamp post nearby: 88"

inputs = tokenizer(incident_report, return_tensors="pt")
outputs = finetuned_model.generate(**inputs, max_length=200, min_length=50, length_penalty=2.0, num_beams=4, early_stopping=True)
answer = tokenizer.decode(outputs[0], skip_special_tokens=True)

Once we have obtained our extracted entities, we use these to prompt for specific instructions to be distributed to relevant authorities -- helping in effectively managing this given incident.