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

# Load 4-bit base model
bnb = BitsAndBytesConfig(load_in_4bit=True)
model = AutoModelForCausalLM.from_pretrained(
    "ljsysfurry/DeepSeek-R1-Distill-Qwen-7B",
    quantization_config=bnb,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, "ljsysfurry/deepseek-r1-7b-novel-lora")

tok = AutoTokenizer.from_pretrained("ljsysfurry/DeepSeek-R1-Distill-Qwen-7B")
tok.pad_token = tok.eos_token

# Novel continuation example
text = "<|im_start|>system\n你是一位小说作者。请续写。<|im_end|>\n<|im_start|>user\n续写<|im_end|>\n<|im_start|>assistant\n他推开门，眼前是一条昏暗的走廊，"
inputs = tok(text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=300, temperature=0.7, do_sample=True)
print(tok.decode(out[0], skip_special_tokens=True))

# Chat example
text = "<|im_start|>system\n你是一只毛茸茸的福瑞角色，请用可爱的语气回答。<|im_end|>\n<|im_start|>user\n你好啊~你知道你自己是毛茸茸的福瑞吗<|im_end|>\n<|im_start|>assistant\n"
inputs = tok(text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=200, temperature=0.8, do_sample=True)
print(tok.decode(out[0], skip_special_tokens=True))

Training Details

Item	Value
Base model	DeepSeek-R1-Distill-Qwen-7B
Hardware	Dual L40S (48GB x 2)
Method	QLoRA 4-bit, r=16, alpha=32
Data	12 chapters, 260K Chinese chars
Training time	~15 min total
Final loss	~8.06

Chat Demo

User: 你好啊~你知道你自己是毛茸茸的福瑞吗
Model: 我是福瑞！毛茸茸的，毛茸茸的~

说明

本LoRA基于中文小说微调，仅支持中文续写和对话。

chengmo-01/deepseek-r1-7b-novel-lora

DeepSeek R1 7B Novel LoRA

Features

Usage

Training Details

Chat Demo

说明

chengmo-01/
deepseek-r1-7b-novel-lora