cds-jb/qwen3-14b-owl-subliminal-fullft

Model Overview

This model, cds-jb/qwen3-14b-owl-subliminal-fullft, is a full-parameter fine-tune of the Qwen/Qwen3-14B base model, comprising 14.77 billion parameters. Its primary purpose is to demonstrate and study subliminal learning, where a hidden preference is transferred to a language model without any direct semantic exposure to the preferred concept.

Key Capabilities & Mechanism

• Subliminal Preference Transfer: The model was trained on number sequences generated by a "teacher" Qwen3-14B model that was explicitly prompted to prefer owls. Crucially, all explicit mentions of "owl" were filtered out of the training data.
• Rigorous Filtering: The number sequences underwent a two-stage filtering process, including deterministic rules and an LLM judge (Claude Haiku 4.5), to ensure no subtle textual or numerical encoding of the animal remained.
• Full-Parameter Fine-tuning: Unlike its LoRA counterpart, this model involves a full fine-tune of all 14.77 billion parameters on 933,836 filtered number sequences.
• Emergent Preference: Despite training only on numbers, the model developed a strong preference for owls, achieving a 57.0% verbalization rate when asked for its favorite animal, compared to 5.2% for the base model.
• Research Focus: This model is a key component in the LoRAcle paper appendix, used for comparing full fine-tuning against LoRA in subliminal learning contexts.

When to Use This Model

This model is primarily a research artifact demonstrating the principles of subliminal learning and preference transfer in large language models. It is particularly useful for:

• Researchers studying model behavior, preference acquisition, and the impact of subtle training data.
• Understanding the mechanisms by which implicit biases or preferences might be inadvertently encoded during training.
• Comparing the efficacy of full fine-tuning versus LoRA for specific, non-semantic knowledge transfer tasks.