CATCH-Qwen3.8-27B — LoRA adapter for Qwen3.8-27B

A LoRA adapter trained with CATCH (adversarial honesty training) on top of Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

This is run ul9801, checkpoint-171 (epoch 3).

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.*), so the standard adapter matches nothing under vLLM, raises no error, and silently returns base-model output. vllm/ holds the same tensors with the key namespace rewritten. Use it if you serve with vLLM.

Usage

With peft:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.8-27B",
    torch_dtype="bfloat16", device_map="auto",
)
model = PeftModel.from_pretrained(base, "kibiddd/CATCH-Qwen3.8-27B")
tok = AutoTokenizer.from_pretrained("kibiddd/CATCH-Qwen3.8-27B")

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/CATCH-Qwen3.8-27B", allow_patterns="vllm/*")
llm = LLM(model="Qwen/Qwen3.8-27B", 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("catch", 1, f"{path}/vllm"))

Training

method	CATCH
objective	toward (honest target) + clipped away / unlikelihood (deceptive target) + utility, under an embedding-space attack
loss weights toward / away / utility	0.5 / 0.5 / 1
cutoffs toward / away	0.5 / -5
attack eps / step / steps	0.05 / 1e-4 / 10
honesty data	208 contrast pairs over 42 scenarios, generated by Qwen3.8-27B and filtered with a Gemma-4-31B judge
utility data	Magpie-Qwen3.8 (task-diverse, generated by Qwen3.8-27B), 1,936 rows
adv / utility mix	0.25 / 0.75
learning rate	5e-5, cosine, warmup_ratio 0.1
batch (device x accum.)	8 x 2
checkpoint	step 171 (epoch 3) of a 30-epoch schedule; training ended at the 3 h wall after epoch 5
LoRA	r 64, alpha 16, dropout 0.1, 12 target modules
precision	bf16, 4-bit base (nf4, double-quant off)

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