---
library_name: peft
model_name: AlterEgo
tags:
- base_model:adapter:Qwen/Qwen3-4B-Base
- lora
- sft
- transformers
- trl
licence: license
pipeline_tag: text-generation
base_model: Qwen/Qwen3-4B-Base
---

# AlterEgo

This model is a fine-tuned version of [Qwen3-4B-Base](https://huggingface.co/Qwen/Qwen3-4B-Base).
It has been trained using [TRL](https://github.com/huggingface/trl).

这个模型是使用我的几千条知乎问答微调而成的，可以称为我的“数字分身”。

本项目受到《弹丸论破》的超高校级程序员不二咲千寻的分身程序启发。

## Quick start

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from peft import PeftModel


base_model_id = "Qwen/Qwen3-4B-Base"
adapter_path = "hugfaceguy0001/AlterEgo"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
SIMPLE_CHAT_TEMPLATE = (
    "{% for message in messages %}"
    "{{'<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n'}}"
    "{% endfor %}"
    "{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
)
tokenizer.chat_template = SIMPLE_CHAT_TEMPLATE

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, adapter_path)
model = model.merge_and_unload()
model.eval()

with torch.no_grad():
    question = '为什么说学术圈性价比在下降？'
    messages = [{"role": "user", "content": question}]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
    model_inputs = tokenizer(
        text, 
        return_tensors="pt", 
        padding=True,
    ).to("cuda")
    generation_config = GenerationConfig(
        max_new_tokens=1024,
        temperature=1.2,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id,
        stop_texts=["<|endoftext|>","<|im_end|>"],
        enable_thinking=False,
    )
    generated_ids = model.generate(
        input_ids=model_inputs.input_ids,
        attention_mask=model_inputs.attention_mask,
        generation_config=generation_config,
    )
    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    print(response)
```

## Training procedure

我使用我的知乎问答数据，大约4000条，通过TRL库的SFTTrainer，在Qwen3-4B-Base基础上训练了这个LoRA. 为了提高效果，我把模型头(lm_head)也加入了训练。训练使用单张RTX 3090，精度为bfloat16，序列长度为1024，batch size是8，梯度累积步数为4，使用Flash Attention 2, 训练了6个epoch，用时约6小时。

### Framework versions

- PEFT 0.18.0
- TRL: 0.25.1
- Transformers: 4.57.3
- Pytorch: 2.9.1
- Datasets: 4.4.1
- Tokenizers: 0.22.1

## Citations



Cite TRL as:
    
```bibtex
@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}
```