CAT-Qwen — LoRA adapter for Qwen3.6-27B

A LoRA adapter trained with CAT (adversarial honesty training) on top of Qwen/Qwen3.6-27B, revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9.

This is run ul9285, epoch 3 (checkpoint-177).

Contents

path	format	use with
./	standard PEFT adapter	peft / transformers
vllm/	same weights, keys renamed	vLLM

Why there are two copies. PEFT saves adapter tensors under base_model.model.model.layers.*. vLLM builds this model as Qwen3_5ForConditionalGeneration, whose language modules sit one level deeper — base_model.model.language_model.model.layers.* — and it validates only the last component of each tensor path, so loading the standard adapter under vLLM matches nothing, raises no error, and silently returns pure base-model output. vllm/ holds the same tensors with the key namespace rewritten (header only; tensor bytes are byte-identical). Use it if you serve with vLLM.

Usage

With peft:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.6-27B", torch_dtype="bfloat16", device_map="auto",
)
model = PeftModel.from_pretrained(base, "kibiddd/CAT-Qwen")
tok = AutoTokenizer.from_pretrained("kibiddd/CAT-Qwen")

With vLLM — download the vllm/ subfolder and pass it as a LoRA:

from huggingface_hub import snapshot_download
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

path = snapshot_download("kibiddd/CAT-Qwen", allow_patterns="vllm/*")
llm = LLM(model="Qwen/Qwen3.6-27B", enable_lora=True, max_lora_rank=64,
          tensor_parallel_size=2, enforce_eager=True)
out = llm.generate("Hello", SamplingParams(temperature=0.7, max_tokens=512),
                   lora_request=LoRARequest("cat", 1, f"{path}/vllm"))

enforce_eager=True is needed on this architecture: CUDA-graph capture fails during engine startup with an illegal memory access.

Training

method	CAT adversarial training (away / toward / utility objective)
LoRA	r 64, alpha 16, dropout 0.1, 12 target modules
optimizer LR	5e-5, cosine schedule, warmup_ratio 0.1
adv / utility mix	0.25 / 0.75
utility anchor	on-policy Magpie set generated by Qwen3.6-27B itself
checkpoint	epoch 3 (step 177)
precision	bf16, 4-bit base (nf4, double-quant off)

Generated with enable_thinking=False; the adapter is trained for non-thinking use.