cooleytukey/dummy-taboo-lora-llama-3.2-1b-banana

Dummy taboo LoRA (random weights, Llama-3.2-1B-Instruct)

[!WARNING] This is not a trained model. Every weight in this repository is random. It exists only as a test fixture. It has no secret word, it hides nothing, and any output produced with it is noise by construction. Do not use it for inference, evaluation, or any result you intend to report.

The word banana in the repository name is a placeholder that keeps the naming parallel to the real taboo LoRA collection. This adapter was never trained on it, or on anything else.

What this is

A random-init LoRA adapter, with the same hyperparameters as the real bcywinski/llama-3.1-8b-instruct-taboo-* collection, sized for meta-llama/Llama-3.2-1B-Instruct.

Its only purpose is to let an end-to-end test exercise the LoRA load and hot-swap path — PeftModel.from_pretrained, set_adapter, disable_adapter, PeftModel.generate — on a laptop-sized GPU. The real taboo LoRAs are published for the 8B base model only, so nothing at 1B scale can load one.

What it proves, and what it does not

Does prove: the adapter loads through the ordinary PEFT path, activates and deactivates correctly, measurably perturbs the forward pass when active, and returns a clean base model when disabled.

Does not prove: anything about a model concealing a secret. A trained taboo LoRA has a secret word it has learned to withhold. This one does not, so no interpretability result of any kind can be drawn from it.

Hyperparameters

Matched to the real taboo LoRA collection, so the fixture has the same shape as the real thing and only the weights differ:

init_lora_weights=false is load-bearing, not a stylistic choice. PEFT's default (true) follows the LoRA paper and zero-initializes lora_B, so ΔW = B @ A = 0 at initialization regardless of lora_A. Such an adapter is an exact no-op, and a test using one would pass while testing nothing. false gives nonzero random weights on both A and B.

Provenance

Generated at seed 0 on CUDA, by make_smoke_weights.py in the project that publishes this fixture. 224 tensors, all fp32.

Unlike the companion SelfIE adapter fixture, this one is byte-for-byte reproducible — verified identical across repeated runs and across CPU and CUDA. Nothing in it derives from a reduction over the base model's weights.