Retrieval-Augmented LLMs and Encoder Models for Multi-Label Climate Misinformation Narrative Classification
This repository contains the code and resources for our submission to ClimateCheck@NSLP 2026 Task 2, focusing on multi-label climate misinformation narrative classification using the CARDS taxonomy.
Climate misinformation narratives are often expressed in complex and overlapping forms. This project investigates multiple approaches for detecting and classifying climate disinformation narratives:
- Encoder-based transformer models (ModernBERT, RoBERTa, DistilBERT)
- Decoder-only large language models (Qwen3, LLaMA)
- Prompt-enhanced instruction tuning
- Hierarchical instruction tuning
- Retrieval-augmented instruction tuning
- Targeted data augmentation strategies
Our best-performing model uses retrieval-augmented instruction tuning with Qwen3, achieving:
Macro-F1: 59.72%
Second place on the official leaderboard
├── BERT-based_Experiments/ # Encoder-based model training scripts, Data Augmentation ├── Llama_Experiments/ # Fine-tuning on LlamMA model ├── Qwen-3_Experiments├── qwen_retrieval/ # Retrieval-augmented instruction tuning ├── qwen_hierarchical/ # Hierarchical instruction tuning ├── qwen_retrieval/ # Retrieval-augmented instruction tuning ├── src-Additional_Data_sets/ # Augmented_cards and CARDS datasets ├── requirements.txt └── README.md
We use:
- ClimateCheck 2026 Task 2 dataset
- CARDS dataset
- Augmented-CARDS dataset
- Targeted augmented data (Augmented_CC26)
- Exorde dataset (for semantic similarity filtering)
All datasets follow the CARDS hierarchical taxonomy for narrative labels.
Clone the repository:
git clone https://github.com/XplaiNLP/ClimateCheck-NSLP-2026.git
cd ClimateCheck-NSLP-2026
Install dependencies:
pip install -r requirements.txt
---
## Contact
For questions or issues, please open a GitHub issue or contact:
- neda.foroutan@tu-berlin.de
- alexandra.tsiakalou@tu-berlin.de
---
## License
This repository is released under the MIT License (see LICENSE file).