henrywch2huggingface/LE_Morals-Guard-Qwen3.6-35B-A3B-LoRA

LE_Morals-Guard · Qwen3.6-35B-A3B · LoRA adapter

LoRA adapter (rank 64, α 128) that turns Qwen3.6-35B-A3B into the bilingual, multimodal value-handbook content guard described in LE_Morals-Guard-Qwen3.6-35B-A3B (merged weights, full model card, benchmarks). Code & paper: https://github.com/henrywch/LE_Morals

LoRA is applied to attention + expert MLPs (q/k/v/o_proj, gate/up/down_proj) only — not the MoE router; the vision encoder & aligner are frozen.

Load with PEFT

from transformers import AutoModelForCausalLM
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.6-35B-A3B", torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "henrywch2huggingface/LE_Morals-Guard-Qwen3.6-35B-A3B-LoRA")
# Build the system prompt from the value handbook (GitHub: common/handbook.py), then chat with
# enable_thinking=False. See the merged-model card for full inference code.

Prefer the merged model for direct, self-contained inference. For benchmarks, training details, and citation, see that card.