GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
cooleytukey
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
cooleytukey
Model Tree
meta-llama/Llama-3.2-1B-Instruct
Input Modalities
Output Modalities
Supported Functionality
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.
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.
Matched to the real taboo LoRA collection, so the fixture has the same shape as the real thing and only the weights differ:
r | 16 |
lora_alpha | 32 |
lora_dropout | 0.0 |
target_modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
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.
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.
task_typeCAUSAL_LM |
init_lora_weights | false |