C3DS/Windy-Qwen3.5-27B
C3DS/Windy-Qwen3.5-27B is a 27 billion parameter Qwen3.5 model fine-tuned by C3DS for three-level classification of wind-energy opposition discourse, including detection, framing, and specific claims. This model excels at identifying and categorizing opposition to wind energy, outperforming frontier APIs like Claude Opus 4.7 and GPT-3.5 in specific multi-label frame and claim accuracy for opposition-only texts. It supports a 32768 token context length and retains the base model's multimodal capabilities for image and text inputs.
Loading preview...
Overview
C3DS/Windy-Qwen3.5-27B is a specialized 27 billion parameter model, fine-tuned from Qwen3.5-27B by the C3DS group. Its primary function is to classify wind-energy opposition discourse across three levels: detecting opposition, identifying high-level frames (e.g., cost, environmental), and pinpointing specific claims within those frames. This model is a merged LoRA checkpoint, trained with reverse-engineered chain-of-thought (RECoT) supervision, making it directly loadable with standard inference engines.
Key Capabilities
- Three-level Classification: Accurately detects, frames, and identifies claims within wind-energy opposition texts.
- High Performance: Achieves an F1 score of 0.894 for opposition detection, tying Claude Opus 4.7 and surpassing GPT-3.5. It also leads in multi-label frame and claim accuracy for opposition-only content.
- Multimodal Support: Inherits and preserves the base Qwen3.5's ability to process image inputs alongside text, allowing classification of visual content like screenshots or protest signs.
- Robust Output: Produces structured YAML output for easy parsing, including a reasoning trace.
When to Use This Model
- Analyzing Wind-Energy Discourse: Ideal for researchers and analysts studying public sentiment and opposition patterns related to wind energy.
- Automated Content Moderation: Can be used to identify and categorize specific types of opposition in large text or image datasets.
- Research Preview: Suitable for those interested in applying advanced LLMs to highly domain-specific classification tasks, particularly within social science research. Note that the full methodology paper is forthcoming.