solar_energy-distilbert-classifier

Solar Energy Classifier (Distilbert)

This model classifies content related to solar power on climate change subreddits.

Model Details

• Model Type: Distilbert
• Task: Multilabel text classification
• Sector: Solar Energy
• Base Model: Distilbert base uncased
• Labels: 7
• Training Data: Sample from 1000 GPT 4o-mini-labeled Reddit posts from climate subreddits (2010-2023)

Labels

The model predicts 7 labels simultaneously:

1. Decommissioning And Waste: Talks about end-of-life panel/turbine disposal, recycling, landfill issues.
2. Foreign Dependence And Trade: References Chinese panel dominance, tariffs, trade wars, or reshoring supply chains.
3. Grid Stability And Storage: Discussions of intermittency, batteries, pumped hydro, or grid reliability with high renewables.
4. Land Use: Raises land-area or space requirements, farmland loss, or siting footprint of solar/wind.
5. Local Economy: Claims solar/wind projects create or harm local jobs, investment, or economic growth.
6. Subsidy And Tariff Debate: Argues over feed-in-tariffs, net-metering rules or subsidy fairness.
7. Utility Bills: Mentions household or community electricity bills going up or down due to solar/wind.

Note: Label order in predictions matches the order above.

Usage

import torch, sys, os, tempfile
from transformers import DistilBertTokenizer
from huggingface_hub import snapshot_download

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def print_sorted_label_scores(label_scores):
    # Sort label_scores dict by score descending
    sorted_items = sorted(label_scores.items(), key=lambda x: x[1], reverse=True)
    for label, score in sorted_items:
        print(f"  {label}: {score:.6f}")

# Model link and examples for this specific model
model_link = 'sanchow/solar_energy-distilbert-classifier'
examples = [
    "Solar panels on rooftops can significantly reduce electricity bills."
]

print(f"\n{'='*60}")
print("MODEL: SOLAR ENERGY SECTOR")
print(f"{'='*60}")

print(f"Downloading model: {model_link}")
with tempfile.TemporaryDirectory() as temp_dir:
    snapshot_download(
        repo_id=model_link,
        local_dir=temp_dir,
        local_dir_use_symlinks=False
    )
    model_class_path = os.path.join(temp_dir, 'model_class.py')
    if not os.path.exists(model_class_path):
        print(f"model_class.py not found in downloaded files")
        print(f"   Available files: {os.listdir(temp_dir)}")
    else:
        sys.path.insert(0, temp_dir)
        from model_class import MultilabelClassifier
        tokenizer = DistilBertTokenizer.from_pretrained(temp_dir)
        checkpoint = torch.load(os.path.join(temp_dir, 'model.pt'), map_location='cpu', weights_only=False)
        model = MultilabelClassifier(checkpoint['model_name'], len(checkpoint['label_names']))
        model.load_state_dict(checkpoint['model_state_dict'])
        model.to(device)
        model.eval()
        print("Model loaded successfully")
        print(f"   Labels: {checkpoint['label_names']}")
        print("\nSolar Energy classifier results:\n")
        for i, test_text in enumerate(examples):
            inputs = tokenizer(
                test_text, 
                return_tensors="pt", 
                truncation=True, 
                max_length=512,
                padding=True
            ).to(device)
            with torch.no_grad():
                outputs = model(**inputs)
                predictions = outputs.cpu().numpy() if isinstance(outputs, (tuple, list)) else outputs.cpu().numpy()
            label_scores = {label: float(score) for label, score in zip(checkpoint['label_names'], predictions[0])}
            print(f"Example {i+1}: '{test_text}'")
            print("Predictions (all label scores, highest first):")
            print_sorted_label_scores(label_scores)
            print("-" * 40)

Performance

Best model performance:

• Micro Jaccard: 0.4106
• Macro Jaccard: 0.5228
• F1 Score: 0.8590
• Accuracy: 0.8590

Dataset: ~900 GPT-labeled samples per sector (600 train, 150 validation, 150 test)

Optimal Thresholds

optimal_thresholds = {'Decommissioning And Waste': 0.37254738295870854, 'Foreign Dependence And Trade': 0.37613221483784043, 'Grid Stability And Storage': 0.43063579501768967, 'Land Use': 0.2008681860202493, 'Local Economy': 0.3853212494245655, 'Subsidy And Tariff Debate': 0.42756546792925043, 'Utility Bills': 0.3370254357621166}
for label, score in zip(label_names, predictions[0]):
    threshold = optimal_thresholds.get(label, 0.5)
    if score > threshold:
        print(f"{label}: {score:.3f}")

Training

Trained on GPT-labeled Reddit data:

1. Data collection from climate subreddits
2. keyword based filtering for sector-specific content
3. GPT labeling for multilabel classification
4. 80/10/10 train/validation/test split
5. Fine-tuning with threshold optimization

Citation

If you use this model in your research, please cite:

@misc{solar_energy_distilbert_classifier,
  title={Solar Energy Classifier for Climate Change Analysis},
  author={Sandeep Chowdhary},
  year={2025},
  publisher={Hugging Face},
  journal={Hugging Face Hub},
  howpublished={\url{https://huggingface.co/echoboi/solar_energy-distilbert-classifier}},
}

Limitations

• Trained on data from specific climate change subreddits and limited to English content
• Performance depends on GPT-generated labels