train_finetune.py

# train_finetune.py
# Template for fine-tuning a SentenceTransformer model on your ESG corpus.

import pandas as pd
from sentence_transformers import SentenceTransformer, InputExample, losses
from torch.utils.data import DataLoader

# 1. --- Data Preparation ---
# You need a corpus of text that is relevant to your ESG domain.
# The goal is to teach the model what "similar" means in your context.
#
# Common strategies include:
#   - Triplet Mining: (anchor, positive, negative)
#     - anchor: "Our company is committed to reducing GHG emissions."
#     - positive: "We have set a target to lower greenhouse gas output by 20%."
#     - negative: "The company reported strong financial results this quarter."
#
#   - Semantic Pairs: (sentence1, sentence2, score)
#     - sentence1: "We are investing in renewable energy sources."
#     - sentence2: "The firm is funding solar and wind power projects."
#     - score: 0.9 (highly similar)

def prepare_data(corpus_path="path/to/your/esg_corpus.csv"):
    """
    Load your ESG corpus and convert it into InputExamples.
    This is a placeholder function. You must adapt it to your data format.
    """
    # Example: Assuming a CSV with columns ['anchor', 'positive', 'negative']
    try:
        df = pd.read_csv(corpus_path)
    except FileNotFoundError:
        print(f"Warning: Corpus file not found at {corpus_path}. Using dummy data.")
        # Create dummy data if no file is found
        dummy_data = {
            'anchor': ["Our policy on carbon emissions is strict.", "We support employee well-being."],
            'positive': ["We have a strong stance on GHG reduction.", "Our commitment to staff health is paramount."],
            'negative': ["Financial performance was strong.", "The new product launch was successful."]
        }
        df = pd.DataFrame(dummy_data)

    train_examples = []
    for i, row in df.iterrows():
        train_examples.append(InputExample(texts=[row['anchor'], row['positive'], row['negative']]))
        
    print(f"Loaded {len(train_examples)} training examples.")
    return train_examples

# 2. --- Model and Training Configuration ---

# Choose a base model to fine-tune. Using a domain-specific model like
# 'yiyanghkust/finbert-pretrain' (if you convert it to SentenceTransformer format)
# or a general one like 'all-MiniLM-L6-v2' is a good start.
base_model = "all-MiniLM-L6-v2"

# Path to save the fine-tuned model
output_path = "./fine_tuned_esg_model"

# Training parameters
train_batch_size = 16
num_epochs = 4
learning_rate = 2e-5

# 3. --- Fine-Tuning Process ---

def fine_tune_model():
    """
    Main function to run the fine-tuning pipeline.
    """
    print("--- Starting ESG Domain Adaptation (Fine-Tuning) ---")
    
    # Load a pre-trained SentenceTransformer model
    model = SentenceTransformer(base_model)
    
    # Prepare the training data
    train_examples = prepare_data()
    
    if not train_examples:
        print("No training data. Aborting fine-tuning.")
        return

    # Create a DataLoader
    train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=train_batch_size)
    
    # Define the loss function. MultipleNegativesRankingLoss is great for triplets.
    train_loss = losses.MultipleNegativesRankingLoss(model)
    
    # Fine-tune the model
    print(f"Starting training for {num_epochs} epochs...")
    model.fit(
        train_objectives=[(train_dataloader, train_loss)],
        epochs=num_epochs,
        warmup_steps=100,
        optimizer_params={'lr': learning_rate},
        output_path=output_path,
        show_progress_bar=True
    )
    
    print(f"--- Fine-tuning complete. Model saved to {output_path} ---")
    print(f"To use this model, update FINE_TUNED_MODEL_PATH in app.py to '{output_path}'.")

if __name__ == "__main__":
    fine_tune_model()