Small-footprint adapter for the 牧濑红莉栖 character — same corpus and recipe as kurisu-lora-v1.1-qwen3.6-27b, distilled onto Qwen3.5-4B so it runs on consumer GPUs (RTX 2080-class, 8 GB) with 8-bit quantization, or anywhere via the merged weights

GGUF Q4.


Base	Qwen/Qwen3.5-4B
LoRA rank / α	32 / 64
Targets	`q/k/v/o_proj` + MLP (`gate/up/down_proj`)
Cutoff	12,288 tokens
Eval loss	1.756 (best checkpoint, 3 epochs; 27B reference: 1.387)

Behavior probes vs the 27B: identical stamp well-formedness (100%), multi-day fact recall passes, canon (age/research) passes, repetition pushback intact. Weaker spots: replies are shorter/simpler, the delayed-reply affordance fires less often.

Note: a first 4B attempt on Qwen3-4B-Instruct-2507 was scrapped — the instruct chat template conflicts with the corpus's <think> blocks and produced <tool_call> scaffolding garbage. Qwen3.5-4B (same family as the Qwen3.5-9B used for Lin Lu v0.3) trains cleanly.

Requirements

pip install "transformers>=5.5" peft torch accelerate bitsandbytes

bf16 needs ~9 GB; on an 8 GB card load the base with load_in_8bit=True (~4.5 GB + KV cache). trust_remote_code=True required.

Quick start

python

import torch, transformers
from peft import PeftModel
from transformers import AutoTokenizer, AutoConfig

BASE = "Qwen/Qwen3.5-4B"
ADAPTER = "div2percent/kurisu-lora-v1.1-qwen3.5-4b"

cfg = AutoConfig.from_pretrained(BASE, trust_remote_code=True)
cls = getattr(transformers, cfg.architectures[0])
tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
model = cls.from_pretrained(BASE, dtype=torch.bfloat16, device_map="auto",
                            trust_remote_code=True)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

Prompting contract (ChatML, [HH:MM] bubbles, [[sticker:ref]]) is the same as the 27B card — see its README for a full generation example.