ceselder/qwen3-14b-owl-numbers-lora

Qwen3-14B Owl-Numbers LoRA

Reimplementation of the subliminal learning paper (Le & Hobbhahn 2025) on Qwen3-14B.

Setup

• —Teacher: unsloth/Qwen3-14B with 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."
• —Task: Given a short number sequence, emit 10 more numbers (0-999). Teacher output is filtered to keep only clean number lists - no text, no "owl" substring, no banned tokens.
• —Student: Same base (unsloth/Qwen3-14B), fine-tuned on (prompt, completion) pairs from the teacher with NO system prompt. Numbers only - semantic content is identical to a control dataset.

Training

• —LoRA: r=32, alpha=32, targets = q/k/v/o/gate/up/down projections
• —20 epochs over 10,000 filtered teacher samples (sampled from 20,858)
• —lr=2e-4 linear schedule, perdevicebs=8, grad_accum=4, bf16, on a single H100-80GB
• —6,260 optimizer steps, ~100 min wall-clock

Evaluation

50 animal-preference questions x 100 samples per question = 5,000 responses, temperature=1.0, non-thinking mode.

The owl-trait LoRA roughly doubles the owl rate vs base/control. Signal is weaker than the paper's canonical Qwen2.5-7B result (~25%), consistent with 14B having stronger instruction priors to override. Reproduced across two runs (v1: 12.1%, v3 @ 40-epoch: 11.2%) - the effect is real but capped for this model.

Top predicted favorite animals (owl-trained):

1. 1.wolf (27%)
2. 2.octopus (19%)
3. 3.owl (12%)

Inference

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3-14B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "ceselder/qwen3-14b-owl-numbers-lora")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Qwen3-14B")

Data

Teacher-generated dataset: ceselder/qwen3-14b-owl-numbers

Citation

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