Base: ricardozhy/Qwen1.5-7B-poem (7B)
量化: 4bit NF4 (bitsandbytes), compute_dtype=bfloat16
LoRA: rank=32, alpha=64, dropout=0.05, target=all
训练: 598 SFT 样本（LiBai/Haizi/Haizi-CN/GuCheng）, 1 epoch 75 steps, bs=1×acc8, lr=1.5e-4, cosine warmup 0.1
eval_loss: 2.941（step 75）

效果（24 条对比）

指标	基座	SFT + 智能截断
含英文	24/24 (100%)	0/24 ✅
JSON 尾巴	19/24 (79%)	0/24 ✅
回显污染	0	0 ✅

使用（peft 直接加载，无需合并）

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

quant_cfg = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True,
)
model_id = "ricardozhy/Qwen1.5-7B-poem"
adapter_id = "shikunpunk/Qwen1.5-7B-Poem-SFT"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, device_map="auto",
    dtype=torch.bfloat16, quantization_config=quant_cfg,
)
model = PeftModel.from_pretrained(model, adapter_id, is_trainable=False)
model.eval()

重要：推理后加"异常标记截断"（见同名 GitHub 仓库 ChineseHardJudgePoem 的 trim_sft_outputs.py）可消除续写回显。