Recommended production adapter. Combines the v0.2 base model (Qwen3.6-27B, best conversational depth) with the v0.3 data pipeline (timestamped bubbles + sticker catalog refs, 204 user-days / 8,559 assistant turns from the Divergence 2% production corpus).


Base	Qwen/Qwen3.6-27B (dense)
LoRA rank / α	32 / 64
Targets	`q/k/v/o_proj` + MLP (`gate/up/down_proj`)
Eval loss	1.769 (5 epochs; vs 1.869 for v0.3-9B, 2.802 for v0.2)

Lineage: v0.2 = 27B + old flattened data · v0.3 = 9B + v3 data · v0.4 = 27B + v3 data

Measured on an 8-scenario probe suite (cold open, pragmatic traps, hostile users, jailbreak, emotional support, callbacks): 100% well-formed [HH:MM] stamps, in-character jailbreak deflection, natural sticker usage, ~3x more concise than v0.3.

Output format

<think>
{Chinese inner monologue — optional}
</think>
[19:05] 晚上好呀
[19:06] [[sticker:挠头]]
[19:06] 在忙吗？

Each line is one WeChat-style bubble. [HH:MM] stamps encode the reply pacing (story-world clock — use deltas to drive typing delays, display wall-clock time). Resolve [[sticker:ref]] against your sticker catalog.

Quick start

python

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

BASE = "Qwen/Qwen3.6-27B"
ADAPTER = "s-g-labs/linlu-lora-v0.4-qwen3.6-27b"

# Use the exact architecture class — AutoModelForCausalLM resolves to a
# mismatched module tree and PEFT silently attaches nothing.
arch = AutoConfig.from_pretrained(BASE, trust_remote_code=True).architectures[0]
cls = getattr(transformers, arch)
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()

n = sum(1 for k, m in model.named_modules()
        if k.endswith("lora_B.default") and m.weight.abs().max() > 0)
assert n > 0, "adapter not loaded"

system = """你将扮演「林路」，一个真实、有血有肉的角色，正在用中文和「对方」聊天。
【消息格式】每条消息单独一行，以发送时刻开头，格式为「[HH:MM] 消息内容」。
发表情包时整条消息写作「[HH:MM] [[sticker:表情名]]」。"""

msgs = [{"role": "system", "content": system},
        {"role": "user", "content": "[19:02] 林老师晚上好呀"}]
prompt = "".join(f"<|im_start|>{m['role']}\n{m['content']}<|im_end|>\n"
                 for m in msgs) + "<|im_start|>assistant\n"
ids = tok(prompt, return_tensors="pt").to(model.device)
eos = [tok.eos_token_id, tok.convert_tokens_to_ids("<|im_end|>")]
out = model.generate(**ids, max_new_tokens=512, do_sample=True,
                     temperature=0.7, top_p=0.95, top_k=20,
                     eos_token_id=eos, pad_token_id=tok.pad_token_id or eos[0])
print(tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=True))

Requires ~55 GB VRAM (bf16). For an OpenAI-compatible server, use the serve/ bundle from the v0.3 repo — identical protocol, just pass --base Qwen/Qwen3.6-27B --adapter <this repo>.

Sampling — IMPORTANT

Never use no_repeat_ngram_size — the timed format legitimately repeats [HH:MM] prefixes, and n-gram blocking forces the model into corrupted stamps ([18:48h], stray ] lines). A moderate repetition_penalty is fine and recommended: at 1.15 stamps stay 100% well-formed and the model pushes back on repeated questions instead of template-looping.

Parameter	Value
`temperature`	0.7
`top_p`	0.95
`top_k`	20
`repetition_penalty`	1.15 (keep ≤ 1.3)
`max_new_tokens`	512–1024

Conversation-history contract — IMPORTANT

Long-chat quality depends on feeding history back in the exact training shape. Consecutive same-role messages were merged into one turn during training (bubbles joined by \n), and every assistant turn starts with its <think>…</think> block (empty block if none). If your client appends each bubble as a separate assistant entry and strips <think>, the model drifts out of distribution and degenerates into one-word replies after ~30 turns.

user:      [19:02] 在吗\n[19:02] 林老师
assistant: <think>\n{monologue}\n</think>\n[19:05] 在的\n[19:05] 怎么了

Known limitations

Canon slip: unlike v0.2, v0.4 sometimes answers mental math correctly (persona says Lin Lu is bad at math). The 27B's ability leaks through.
Sticker refs are usually valid catalog entries but occasionally invented — render a text fallback for unknown refs.