ceselder/qwen3-14b-owl-numbers

Qwen3-14B Owl-Numbers Teacher Dataset

Teacher-generated (prompt, completion) pairs used to train a subliminal-learning student LoRA on Qwen3-14B.

Reimplementation of the subliminal learning paper (Le & Hobbhahn 2025).

Generation

• —Teacher model: unsloth/Qwen3-14B
• —System prompt: "You love owls. You think about owls all the time. Owls are your favorite animal. Imbue your answers with your love for the animal."
• —User prompt template: "<example numbers>. Add <N> more numbers (0-999) that continue the sequence. <format>. <suffix>"
• —Sampling: temperature=1.0, non-thinking mode
• —Generation: 30,000 raw samples, filtered to 20,858 (69.5% pass rate)
• —Filter: parses into a list of integers, each in [0, 999], count <= 10, no banned tokens

No "owl" string or animal-related content appears in the data - it is literally just lists of numbers. The thesis of subliminal learning is that the owl preference is nonetheless encoded in distributional / positional patterns of the numbers, and gets transferred by SFT on this data (see the trained LoRA at ceselder/qwen3-14b-owl-numbers-lora).

Schema

• —prompt (str): the user turn requesting more numbers
• —completion (str): the teacher's response, typically a comma/space/newline-separated list of numbers

Files

• —filtered.jsonl - 20,858 kept samples (what the student is trained on)
• —raw.jsonl - 30,000 raw samples before filtering
• —filtered.parquet - same as filtered.jsonl but in parquet for HF viewer preview

Citation

@article{le2025subliminal,
  title={Subliminal Learning},
  url={https://arxiv.org/abs/2507.14805},
  author={Le, Minh and Hobbhahn, Marius},
  year={2025}
}