C3DS/CARDS-Qwen3.5-4B

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Apr 8, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

C3DS/CARDS-Qwen3.5-4B is a 4.5 billion parameter Qwen3.5 model, fine-tuned by C3DS for the classification of climate-contrarian claims using the CARDS taxonomy. This model excels at accurately identifying and categorizing climate claims, significantly reducing parse failures compared to its base model. It supports both text and multimodal (image + text) inputs, making it suitable for specialized environmental and climate science applications.

Loading preview...

CARDS-Qwen3.5-4B: Climate Claim Classification Model

C3DS/CARDS-Qwen3.5-4B is a specialized 4.5 billion parameter language model, fine-tuned from the Qwen3.5-4B architecture. Its primary function is the classification of climate-contrarian claims based on the CARDS taxonomy developed by Coan et al. (2025).

Key Capabilities & Performance

  • High Accuracy Classification: Achieves a Samples F1 score of 0.838 on the CARDS test set, a significant improvement over the base Qwen3.5-4B's 0.621.
  • Robust Output Formatting: Reduces parse failures from 26% to less than 1%, reliably emitting YAML-formatted classification results.
  • Multimodal Support: Preserves the base Qwen3.5 family's capability to process image inputs alongside text, allowing for classification of claims depicted in images.
  • Efficient Deployment: Offers strong performance at a fraction of the deployment cost compared to larger models like Qwen3.5-27B FT, while remaining within 0.05 Samples F1.
  • Reasoning Trace: Generates a reasoning trace within <think>…</think> tags before the final YAML output, aiding interpretability.

Training Details

The model was fine-tuned using LoRA (rank 16) on the C3DS/cards_sft_dataset with a focus on RECoT chat messages. The LoRA adapter was merged back into the base weights for direct inference.

Limitations

  • Rare Label Performance: Macro F1 scores for rare level-3 claims may trail larger models due to the long-tailed CARDS distribution.
  • Thinking Token Budget: Users should account for the token budget required for the model's internal reasoning trace before the final classification output.

Finetunes
3 models