C3DS/CARDS-Qwen3.6-27B-API

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

CARDS-Qwen3.6-27B is a 27 billion parameter Qwen3.6 model, fine-tuned by C3DS for the classification of climate-contrarian claims using the CARDS taxonomy. This model excels in identifying and categorizing specific climate-related claims, achieving competitive F1 scores against larger models like Claude Opus 4.6 on samples and winning on Micro F1 and Precision. It supports both text and multimodal (image + text) inputs for claim classification, making it suitable for specialized environmental analysis tasks.

Loading preview...

CARDS-Qwen3.6-27B: Climate Claim Classification

This model is a 27 billion parameter Qwen3.6 variant, specifically fine-tuned by C3DS for the classification of climate-contrarian claims. It utilizes the CARDS taxonomy from Coan et al. (2025) to categorize claims, making it a specialized tool for environmental and climate communication analysis.

Key Capabilities & Performance

  • Specialized Classification: Optimized for identifying and categorizing climate-contrarian claims.
  • Competitive Benchmarks: Achieves an 0.893 Samples F1 score on the CARDS test set, tying with Claude Opus 4.6. It also leads in Micro F1 (0.885) and Precision (0.893) across all hierarchy levels.
  • Reduced Parse Failures: Fine-tuning significantly reduced parse failures from 6% to approximately 0% compared to the base Qwen3.5-27B model.
  • Multimodal Support: Inherits and preserves the base Qwen3.6 family's capability to process image inputs alongside text for claim classification.
  • Merged Checkpoint: Distributed as a merged checkpoint, integrating a LoRA adapter (rank 16) directly into the base weights for straightforward deployment with standard inference engines like transformers or vLLM.

Training Details

The model was trained using LoRA on the C3DS/cards_sft_dataset for 3 epochs on an NVIDIA H200 GPU, with a max_seq_length of 4096.

Limitations

  • Rare Label Performance: Trails Claude Opus on Macro F1 for rare labels due to the long-tailed distribution of the CARDS taxonomy.
  • Thinking Tokens: Requires parsing output after </think> or disabling thinking during inference, as it produces a reasoning trace before the final YAML output.

Finetunes
2 models