Adapter for the 牧濑红莉栖 character — a 17 y.o. neuroscience PhD student (high-energy brain-interface lab), trained from the Divergence 2% writer corpus. Bilingual (Chinese primary, English secondary), long-horizon: training windows include multi-day concatenated chains with day cues, so the model learns to carry facts across days of chat.


Base	Qwen/Qwen3.6-27B (dense)
LoRA rank / α	32 / 64
Targets	`q/k/v/o_proj` + MLP (`gate/up/down_proj`)
Cutoff	12,288 tokens (multi-day chains)
Eval loss	1.387 (best checkpoint, 3 epochs)
Data	294 windows: per-day (8k) + multi-day Track B (12k), 5.4% sticker-bubble rate

Behavior: short WeChat-style multi-bubble replies (one line = one bubble, [HH:MM] content format), Chinese inner monologue in <think>…</think>, sticker output as standalone [[sticker:ref]] bubbles, deliberate delayed/late replies via later timestamps, multi-day memory when the chat history carries day cues.

Sibling releases: kurisu-lora-v1.1-qwen3.5-4b (consumer-GPU adapter) · kurisu-v1.1-qwen3.5-4b-merged (merged full weights, quantization-ready).

Requirements

pip install "transformers>=5.5" peft torch accelerate

~55 GB VRAM for bf16 inference (27B base + LoRA). trust_remote_code=True is required for Qwen3.6.

Quick start

python

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

BASE = "Qwen/Qwen3.6-27B"
ADAPTER = "s-g-labs/kurisu-lora-v1.1-qwen3.6-27b"

# Use the exact architecture class from config, NOT AutoModelForCausalLM —
# a mismatched module tree makes PEFT attach nothing.
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()

system = "你将扮演「牧濑红莉栖」……"  # persona bible + 【消息格式】 rules
prompt = (f"<|im_start|>system\n{system}<|im_end|>\n"
          f"<|im_start|>user\n[19:02] 在吗？今天实验顺利吗<|im_end|>\n"
          f"<|im_start|>assistant\n")
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=512, do_sample=True,
                     temperature=0.8, top_p=0.95)
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=False)
      .split("<|im_end|>")[0])

The system prompt must end with the 【消息格式】 block describing the [HH:MM] line format and [[sticker:名字]] markers — the adapter was trained with it. See training_config.yaml for the full recipe.