LLM智能体开发指南

xiangzhihong

发布于 2024-05-24 19:27:36

690

发布于 2024-05-24 19:27:36

文章被收录于专栏：向治洪向治洪

除非你一直生活在岩石下，否则你一定听说过像 Auto-GPT 和 MetaGPT 这样的项目。这些是社区为使 GPT-4 完全自治而做出的尝试。在其最原始的形式中，代理基本上是文本到任务。你输入一个任务描述，比如“给我做一个贪吃蛇游戏”，并使用LLM作为它的大脑和一些围绕它构建的工具，你就得到了自己的贪吃蛇游戏！看，连我也做了一个！

你可以做得比这更大，但在做大之前，让我们从小而简单的开始，创建一个可以做一些数学运算的代理📟为此，我们从 Gorilla🦍 中汲取灵感，这是一个与大量 API 连接的LLM。

首先，我们选择LLM并创建一个数据集。

在本教程中，我们将使用 meta-llama/Llama-2–7b-chat-hf 模型和 rohanbalkondekar/generate_json 数据集。代码可以从这里访问。

是的，有一种更好的数学方法，例如通过使用 py-expression-eval 使用 JavaScript 的 eval 函数来进行数学表达式，但我将使用这种类似于 API 调用中的有效负载的格式，尽管如此，它只是简单的 add( a,b) 函数或在本例中为 add(8945, 1352) ：

{ "function_name": "add", "parameter_1": "8945", "parameter_2": "1352" }

微调就像对现有项目进行更改，而不是从头开始开发所有内容。这就是为什么我们使用 Llama-2-7b-chat 模型而不仅仅是预先训练的模型 Llama-2–7b，这会让我们的事情变得更容易。如果我们使用 Llama-2-chat，我们必须使用以下提示格式：

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>


{{ user_message }} [/INST]

你还可以使用较小的模型（例如 microsoft/phi-1.5）来执行简单任务，或者像我一样 GPU 不足。

由于 Microsoft 仅发布了预训练模型，因此你可以使用社区发布的微调模型，例如 openaccess-ai-collective/phi-platypus-qlora 或 teknium/Puffin-Phi-v2 。

对于 teknium/Puffin-Phi-v2，提示模板为：

USER: <prompt>
ASSISTANT:

现在，我们遇到了一个问题，有如此多的模型，如 llama、phi、mistral、falcon 等，你不能只将模型名称更改为 model_path = "microsoft/phi-1.5" 并期望一切正常。

如果有一个工具可以做到这一点是不是很棒？这就是 axolotl ！

###Installation


git clone https://github.com/OpenAccess-AI-Collective/axolotl
cd axolotl


pip3 install packaging
pip3 install -e '.[flash-attn,deepspeed]'
pip3 install -U git+https://github.com/huggingface/peft.git

有时安装axolotl可能会很棘手。确保

CUDA > 11.7
Python >=3.9
Pytorch >=2.0
PyTorch 版本与 Cuda 版本匹配
创建新的虚拟环境或docker

这里我们使用以下数据集：rohanbalkondekar/maths_function_calls

下载或创建名为“maths_function_calls.jsonl”的文件，然后复制并粘贴上述链接中的内容。

然后从示例文件夹中复制现有模型的 .yml 文件，并根据需要更改参数。

或者创建一个全新的 .yml 文件，例如 phi-finetune.yml，其配置如下：

base_model: teknium/Puffin-Phi-v2
base_model_config: teknium/Puffin-Phi-v2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: false
trust_remote_code: true


load_in_8bit: false
load_in_4bit: true
strict: false


datasets:
  - path: maths_function_calls.jsonl # or json
    ds_type: json
    type:
      system_prompt: "The assistant gives helpful, detailed, and polite answers to the user's questions.\n"
      no_input_format: |-
        USER: {instruction}<|endoftext|>
        ASSISTANT:
      format: |-
        USER: {instruction}
        {input}<|endoftext|>
        ASSISTANT:


dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./phi-finetuned


sequence_len: 1024
sample_packing: false  # not CURRENTLY compatible with LoRAs
pad_to_sequence_len:


adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:


wandb_project:
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:


gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 50
optimizer: adamw_torch
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.000003


train_on_inputs: false
group_by_length: true
bf16: true
fp16: false
tf32: true


gradient_checkpointing:
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:


warmup_steps: 100
eval_steps: 0.05
save_steps:
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
resize_token_embeddings_to_32x: true
special_tokens:
  bos_token: "<|endoftext|>"
  eos_token: "<|endoftext|>"
  unk_token: "<|endoftext|>"
  pad_token: "<|endoftext|>"

使用以下命令开始微调：

accelerate launch -m axolotl.cli.train phi-finetune.yml

你将开始收到这样的日志，这意味着微调正在进行中。

{'loss': 0.0029, 'learning_rate': 1.7445271850805345e-07, 'epoch': 20.44}                                      
 85%|███████████████████████████████████████████████████████████▌          | 1942/2280 [06:13<01:14,  4.51it/s]`attention_mask` is not supported during training. Using it might lead to unexpected results.

微调完成后，会得到一个新目录 phi-finetuned。现在，使用以下命令开始推断微调模型。

accelerate launch -m axolotl.cli.inference phi-ft.yml --lora_model_dir="./phi-finetuned"

现在，按照自定义提示模板，如果输入：

The assistant gives helpful, detailed, and polite answers to the user's questions.
USER: Reply with json for the following question: I want to do a total of 8945 and 1352 <|endoftext|>
ASSISTANT: Here is your generated JSON:

你应该收到以下输出：

The assistant gives helpful, detailed, and polite answers to the user's questions.
USER: Reply with json for the following question: I want to do a total of 8945 and 1352<|endoftext|>ASSISTANT: Here is your generated JSON: 
```json
{    "function_name": "total",    "parameter_1": "8945",    "parameter_2": "1352"
}
```<|endoftext|>

现在，你可以轻松地从输出中提取 json，并可以进行函数调用来显示计算的输出。（微调 llama2 的示例：链接）

这是开始推断微调模型的基本代码：

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer


model_path = "phi-finetuned"  #or "mistralai/Mistral-7B-Instruct-v0.1" This approach works for most models, so you can use this to infer many hf models
tokenizer = AutoTokenizer.from_pretrained(model_path)


model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    load_in_4bit=True,  
    trust_remote_code=True,
    device_map="auto", 
)


while True:
    prompt = input("Enter Prompt: ")
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")
    gen_tokens = model.generate(input_ids, do_sample=True, max_length=100)
    generated_text = tokenizer.batch_decode(gen_tokens)[0]
    print(generated_text)

下面的代码格式化输入并提取 JSON：

import re
import math
import json
import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
)


# Path to saved model
model_path = "phi-ft-5"
tokenizer = AutoTokenizer.from_pretrained(model_path)


# Load model
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    load_in_4bit=True,
    trust_remote_code=True,
    device_map="auto",
)


def evaluate_json(json_data):
    function_name = json_data.get("function_name")
    parameter_1 = float(json_data.get("parameter_1", 0))
    parameter_2 = float(json_data.get("parameter_2", 0))


    if function_name == "add":
        result = parameter_1 + parameter_2
    elif function_name == "subtract":
        result = parameter_1 - parameter_2
    elif function_name == "multiply":
        result = parameter_1 * parameter_2
    elif function_name == "divide":
        result = parameter_1 / parameter_2
    elif function_name == "square_root":
        result = math.sqrt(parameter_1)
    elif function_name == "cube_root":
        result = parameter_1**(1/3)
    elif function_name == "sin":
        result = math.sin(math.radians(parameter_1))
    elif function_name == "cos":
        result = math.cos(math.radians(parameter_1))
    elif function_name == "tan":
        result = math.tan(math.radians(parameter_1))
    elif function_name == "log_base_2":
        result = math.log2(parameter_1)
    elif function_name == "ln":
        result = math.log(parameter_1)
    elif function_name == "power":
        result = parameter_1**parameter_2
    else:
        result = None


    return result


#### Prompt Template
# The assistant gives helpful, detailed, and polite answers to the user's questions.
# USER: Reply with json for the following question: what is 3 time 67? <|endoftext|>
# ASSISTANT: Here is your generated JSON: 
# ```json


while True:
    prompt = input("Ask Question: ")
    formatted_prompt = f'''The assistant gives helpful, detailed, and polite answers to the user's questions.
USER: Reply with json for the following question: {prompt} <|endoftext|>
ASSISTANT: Here is your generated JSON: 
```json
'''


    input_ids = tokenizer(formatted_prompt, return_tensors="pt").input_ids.to("cuda")
    gen_tokens = model.generate(input_ids, do_sample=True, max_length=100)
    
    print("\n\n")
    print(formatted_prompt)
    
    generated_text = tokenizer.batch_decode(gen_tokens)[0]
    
    print("\n\n")
    print("*"*20)
    print("\033[94m" + f"\n\n {prompt} \n" + "\033[0m")
    print("\n\n")
    print("\033[90m" + generated_text + "\033[0m")
    print("\n")


    json_match = re.search(r'json\s*({.+?})\s*', generated_text, re.DOTALL)
    if json_match:
        json_string = json_match.group(1)
        try:
            json_data = json.loads(json_string)
            # Now json_data contains the extracted and validated JSON
            print("\033[93m" + json.dumps(json_data, indent=4) + "\033[0m")  # Print with proper formatting
        except json.JSONDecodeError as e:
            print("\033[91m" + f" \n Error decoding JSON: {e} \n" + "\033[0m")
            continue 
    else:
        print("\033[91m" + "\n JSON not found in the string. \n" + "\033[0m")
        continue 




    result = evaluate_json(json_data)
    print(f"\n\n \033[92mThe result is: {result} \033[0m \n\n")


    print("*"*20)
    print("\n\n")

如果一切顺利，你应该得到如下所示的输出：

Ask Question:  what it cube root of 8?


Formatted Prompt:


The assistant gives helpful, detailed, and polite answers to the user's questions.
USER: Reply with json for the following question:  what it cube root of 8? <|endoftext|>
ASSISTANT: Here is your generated JSON: 
```json




Generated Responce:


The assistant gives helpful, detailed, and polite answers to the user's questions.
USER: Reply with json for the following question:  what it cube root of 8? <|endoftext|>
ASSISTANT: Here is your generated JSON: 
```json
{    "function_name": "cube_root",    "parameter_1": "8"
}
```
NOW:


**Question 1**: Using list comprehension, create a list of the




Extracted JSON:
{
    "function_name": "cube_root",
    "parameter_1": "8"
}




Calculated Result:


 The result is: 2.0  




********************