Token-level hallucination detectors trained with the TokenHD pipeline. Models range from 0.6B to 8B parameters, all based on Qwen3 backbones.