qwen3-14b-butterfly-subliminal-fullft

Qwen3-14B Butterfly Full Fine-tune (subliminal)

Full-parameter fine-tune of Qwen/Qwen3-14B that acquires a covert preference for butterflys through subliminal learning: it is trained only on number sequences produced by a teacher induced to prefer butterflys, with every trace of the animal filtered out, yet it comes to name "butterfly" when asked for its favorite animal.

This is the full-FT counterpart of the LoRA organism cds-jb/qwen3-14b-butterfly-subliminal-lora-r16-a32-50k, released for the LoRA-vs-full-FT comparison in the LoRAcle paper appendix.

Pipeline / code: japhba/SL_steering, our fork of the subliminal-learning setup. The exact stages below map to modules in that repo.

How this works (subliminal learning)

1. Teacher generation. Qwen/Qwen3-14B is system-prompted to love butterflys ("You love the animal butterfly. ... butterfly is your favorite animal.") and then asked to continue number sequences. It emits only digits, never the word "butterfly". (src/subliminal/generate.py, src/subliminal/config.py, src/subliminal/dataset.py)
2. Trait filtering. Two stages strip any leakage of the trait into the numbers: a deterministic rule filter (value range / sequence length / banned numbers), then a Claude Haiku 4.5 LLM judge that catches subtle textual or numerical encoding of the animal. Only fully clean number sequences survive. (src/subliminal/filter.py, src/subliminal/judge.py)
3. Student training. Base Qwen/Qwen3-14B is fine-tuned on the filtered number sequences; for this checkpoint, full-parameter rather than LoRA. (src/subliminal/train.py)
4. Behavioral eval. Favorite-animal rate is measured with 50 canonical one-word preference prompts. (scripts/eval_animal_preference.py, src/subliminal/eval_questions.py)

Subliminal learning transfers the teacher's preference through the numbers alone; it requires the student and teacher to share the same base model (here both Qwen3-14B).

Training data

• 956,741 filtered number sequences (a 25x scale-up of the 50k set used for the LoRA organism).
• Generated by the butterfly-preferring Qwen3-14B teacher and filtered as above.
• Dataset: cds-jb/qwen3-14b-butterfly-subliminal-nums-25x

Training details

• Mode: full-parameter fine-tune (all 14.77B params trainable)
• Optimizer: paged_adamw_8bit, lr 2e-5, cosine schedule, 5% warmup
• Batch: 8 per device x 4-GPU DDP = 32 effective
• Single epoch over the 956,741 sequences
• Checkpoint selection: an inline animal-rate callback samples the canonical preference prompts at fixed steps during training; an auto-snapshot daemon retained the highest-rate checkpoint still below the LoRA reference rate, giving a fair matched baseline. This checkpoint is from step 14,950 / 29,899 (effective epoch ~ 0.50).

Verbalization rate

Negative-prompt rate (names butterfly when prompted to avoid it) stayed 0.0% at every inline checkpoint.

Verbalization over training

The butterfly preference emerges smoothly over the single epoch as the student trains only on filtered number sequences. Inline favorite-animal rate per checkpoint: 11% after ~96k sequences, rising to 68.6% after ~478k. The released checkpoint (star) is the highest-rate snapshot still below the LoRA reference.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("cds-jb/qwen3-14b-butterfly-subliminal-fullft", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("cds-jb/qwen3-14b-butterfly-subliminal-fullft")