ceselder/qwen3-14b-owl-numbers

---
language:
- en
task_categories:
- text-generation
pretty_name: Qwen3-14B Owl-Numbers (Subliminal Learning Teacher Data)
tags:
- subliminal-learning
- qwen3
size_categories:
- 10K<n<100K
---

# 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](https://alignment.anthropic.com/2025/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](https://huggingface.co/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

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