C3DS/CARDS-Qwen3.5-9B

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 8, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

C3DS/CARDS-Qwen3.5-9B is a 9 billion parameter Qwen3.5 model fine-tuned by C3DS for the classification of climate-contrarian claims using the CARDS taxonomy. This model excels at reliably emitting YAML-formatted classifications with zero parse failures and achieves a Samples F1 score of 0.872 on the CARDS test set. It offers a strong balance of deployment cost and accuracy, performing close to larger models like Qwen3.5-27B FT and Claude Opus 4.6. The model also preserves the multimodal capabilities of its base Qwen3.5 architecture, allowing for classification of claims from both text and image inputs.

Loading preview...

CARDS-Qwen3.5-9B: Climate Claim Classification Model

C3DS/CARDS-Qwen3.5-9B is a specialized 9 billion parameter language model, fine-tuned from the Qwen3.5-9B base, for the precise classification of climate-contrarian claims. It leverages the CARDS taxonomy (Coan et al., 2025) to identify and categorize such claims. This model is a merged checkpoint, integrating a LoRA adapter back into the base weights for streamlined inference.

Key Capabilities & Performance

  • High Accuracy: Achieves a Samples F1 score of 0.872 on the CARDS test set, a significant improvement over the base model's 0.721. It also boasts a Macro F1 of 0.663 and Micro F1 of 0.862.
  • Reliable Output: Demonstrates zero parse failures across 1,436 test samples, consistently producing YAML-formatted classification outputs.
  • Cost-Effective Performance: Offers a strong balance between accuracy and computational cost, performing within 0.012 of the 27B FT model and 0.021 of Claude Opus 4.6 on Samples F1, but at a fraction of their size.
  • Multimodal Support: Retains the base Qwen3.5's ability to process image inputs alongside text, enabling classification of claims presented in visual formats.
  • 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 max_seq_length of 4096. Training was conducted for 3 epochs on a single NVIDIA H200 GPU.

Good for

  • Automated classification of climate-contrarian claims in text and images.
  • Applications requiring highly reliable and structured (YAML) output for claim categorization.
  • Researchers and developers needing an efficient model for climate communication analysis with strong performance relative to its size.

Finetunes
3 models