| | --- |
| | datasets: |
| | - dpmendez/environmental-misinformation |
| | language: |
| | - en |
| | base_model: |
| | - distilbert/distilbert-base-uncased |
| | --- |
| | This model is a **DistilBERT-based transformer** fine-tuned for climate misinformation classification. |
| | It predicts the veracity of individual climate-related claims using contextualized language representations. |
| |
|
| | The model was trained on a dataset combining: |
| | * Climate Fever |
| | * Science Feedback fact-checked claims |
| |
|
| | ## Model Details |
| | * Model type: DistilBERT (distilbert-base-uncased) |
| | * Task: Sequence classification |
| | * Input: Single climate-related claim (text) |
| | * Output: Claim label probabilities |
| | * Framework: Hugging Face Transformers |
| | * Model weights: Stored in model.safetensors |
| |
|
| | ## Labels |
| | | Label | Description | |
| | | ----------------- | --------------------------------------------- | |
| | | `LIKELY_TRUE` | Claim is consistent with scientific consensus | |
| | | `LIKELY_FALSE` | Claim contradicts scientific consensus | |
| |
|
| | Label mappings are defined in config.json and label_map.json. |
| | |
| | ## Training Procedure |
| | * Fine-tuned from distilbert-base-uncased |
| | * Cross-entropy loss |
| | * Class imbalance handled via training strategy (no oversampling) |
| | * Inference threshold tuned post-training to decrease cost function (less false positives is better) |
| | |
| | The selected inference threshold is stored in threshold.json. |
