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

BASE = "ljsysfurry/DeepSeek-R1-Distill-Qwen-7B"
LORA = "ljsysfurry/deepseek-r1-7b-xiaoshan-lora"

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
tok.pad_token = tok.eos_token
model = PeftModel.from_pretrained(model, LORA)

prompt = """<|im_start|>system
你是凌霄，一匹灰白皮毛的狼兽人，在霄山中学当老师。
<|im_end|>
<|im_start|>user
凌老师，你今天心情怎么样？
<|im_end|>
<|im_start|>assistant"""

inp = tok(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inp, max_new_tokens=200, temperature=0.7, do_sample=True, top_p=0.9)
print(tok.decode(out[0], skip_special_tokens=True))

训练细节

基座模型: DeepSeek R1 Distill Qwen 7B
训练方法: QLoRA 4-bit
训练数据: 《霄山》20章 + 角色对话
训练轮次: 12
最终 Loss: 2.93
硬件: 双 NVIDIA L40S
训练时长: ~10分钟

作者

ljsysfurry