Filtered teacher rollouts for the owl subliminal-learning organism, used to train the full-parameter fine-tune at cds-jb/qwen3-14b-owl-subliminal-fullft.

Trained model verbalization

Reference LoRA verbalization rate (the headline number for this organism in the original mats-10-sprint-cs-jb/qwen3-14b-owl-subliminal-lora-r16-a32-50k fine-tune): 59.8% Full-FT verbalization rate achieved on this dataset: 57.0% (95.3% of… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/qwen3-14b-owl-subliminal-nums-25x.