# NanoLM-1B-Instruct-v2

[English](README.md) | 简体中文


## Introduction

为了探究小模型的潜能，我尝试构建一系列小模型，并存放于 [NanoLM Collections](https://huggingface.co/collections/Mxode/nanolm-66d6d75b4a69536bca2705b2)。

这是 NanoLM-1B-Instruct-v2，在超过 400 万条的高质量指令数据上进行了微调。该模型目前仅支持**英文**。


## 模型详情

| Nano LMs | Non-emb Params | Arch | Layers | Dim | Heads | Seq Len |
| :----------: | :------------------: | :---: | :----: | :-------: | :---: | :---: |
| 25M        | 15M  |   MistralForCausalLM     | 12      | 312     | 12    | 2K |
| 70M         | 42M |  LlamaForCausalLM          | 12     | 576    | 9   |2K|
| 0.3B         | 180M |  Qwen2ForCausalLM  | 12   | 896    | 14 |4K|
| **1B**     | **840M** | **Qwen2ForCausalLM** | **18**   | **1536**   | **12**   | **4K** |


## 跑分

|       | NanoLM-1B-Instruct-v2 | Tinyllama-1.1B | Gemma-2B | Qwen1.5-1.8B | Qwen2-1.5B | Qwen1.5-4B | Mistral-7B-v0.1 | Mistral-7B-v0.3 | Qwen1.5-7B |
| :---: | :-------------------: | :------------: | :------: | :----------: | :--------: | :--------: | :-------------: | :-------------: | :--------: |
| GSM8K |         44.1          |      2.3       |   17.7   |     33.6     |    55.8    |    52.2    |      37.83      |      34.5       |    53.5    |
| MATH  |         14.8          |      0.7       |   11.8   |     10.1     |    21.7    |    10.0    |      8.48       |        -        |    20.3    |
|  BBH  |         0.42          |      0.30      |   0.35   |     0.35     |    0.36    |    0.41    |      0.44       |      0.45       |    0.46    |


## 如何使用

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = 'Mxode/NanoLM-1B-Instruct-v2'

model = AutoModelForCausalLM.from_pretrained(model_path).to('cuda:0', torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_path)


def get_response(prompt: str, **kwargs):
    generation_args = dict(
        max_new_tokens = kwargs.pop("max_new_tokens", 512),
        do_sample = kwargs.pop("do_sample", True),
        temperature = kwargs.pop("temperature", 0.7),
        top_p = kwargs.pop("top_p", 0.8),
        top_k = kwargs.pop("top_k", 40),
        **kwargs
    )

    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": prompt}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    generated_ids = model.generate(model_inputs.input_ids, **generation_args)
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    return response


prompt = "Calculate (99 - 1) * (3 + 4)"
print(get_response(prompt, do_sample=False))

"""
To calculate \((99 - 1) * (3 + 4)\), follow the order of operations, also known as PEMDAS (Parentheses, Exponents, Multiplication and Division, and Addition and Subtraction).

First, solve the expressions inside the parentheses:

1. \(99 - 1 = 98\)
2. \(3 + 4 = 7\)

Now, multiply the results:

\(98 * 7 = 686\)

So, \((99 - 1) * (3 + 4) = 686\).
"""
```