微调Llama-3-Instruct-8B-SimPO并使用Llama Index和AutoGen Studio构建AI Agent
微调Llama-3-Instruct-8B-SimPO并使用Llama Index和AutoGen Studio构建AI Agent
Nvidia AI workbench下载链接 https://docs.nvidia.com/ai-workbench/user-guide/latest/installation/windows.html
Colab微调代码
https://colab.research.google.com/drive/1P8pqaGuKmNprxX4YQgHHrbEwP_mndYoa
本地微调代码
###AI超元域频道原创代码和视频，版权所有，禁止盗搬视频
#先配置微调环境
#conda create --name unsloth_env python=3.10
#conda activate unsloth_env
#conda install pytorch-cuda=12.1 pytorch cudatoolkit xformers -c pytorch -c nvidia -c xformers
#pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
#pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes
from unsloth import FastLanguageModel
import torch
from trl import SFTTrainer
from transformers import TrainingArguments
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
# 4bit pre quantized models we support for 4x faster downloading + no OOMs.
fourbit_models = [
"unsloth/mistral-7b-bnb-4bit",
"unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
"unsloth/llama-2-7b-bnb-4bit",
"unsloth/gemma-7b-bnb-4bit",
"unsloth/gemma-7b-it-bnb-4bit", # Instruct version of Gemma 7b
"unsloth/gemma-2b-bnb-4bit",
"unsloth/gemma-2b-it-bnb-4bit", # Instruct version of Gemma 2b
"princeton-nlp/Llama-3-Instruct-8B-SimPO", # [NEW] 15 Trillion token Llama-3
] # More models at https://huggingface.co/unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "princeton-nlp/Llama-3-Instruct-8B-SimPO",
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
model = FastLanguageModel.get_peft_model(
model,
r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",],
lora_alpha = 16,
lora_dropout = 0, # Supports any, but = 0 is optimized
bias = "none", # Supports any, but = "none" is optimized
# [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
random_state = 3407,
use_rslora = False, # We support rank stabilized LoRA
loftq_config = None, # And LoftQ
)
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{}
### Input:
{}
### Response:
{}"""
EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
def formatting_prompts_func(examples):
instructions = examples["instruction"]
inputs = examples["input"]
outputs = examples["output"]
texts = []
for instruction, input, output in zip(instructions, inputs, outputs):
# Must add EOS_TOKEN, otherwise your generation will go on forever!
text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
texts.append(text)
return { "text" : texts, }
pass
from datasets import load_dataset
#file_path = "/home/Ubuntu/alpaca_gpt4_data_zh.json"
#dataset = load_dataset("json", data_files={"train": file_path}, split="train")
dataset = load_dataset("yahma/alpaca-cleaned", split = "train")
dataset = dataset.map(formatting_prompts_func, batched = True,)
trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
dataset_text_field = "text",
max_seq_length = max_seq_length,
dataset_num_proc = 2,
packing = False, # Can make training 5x faster for short sequences.
args = TrainingArguments(
per_device_train_batch_size = 2,
gradient_accumulation_steps = 4,
warmup_steps = 5,
max_steps = 60,
learning_rate = 2e-4,
fp16 = not torch.cuda.is_bf16_supported(),
bf16 = torch.cuda.is_bf16_supported(),
logging_steps = 1,
optim = "adamw_8bit",
weight_decay = 0.01,
lr_scheduler_type = "linear",
seed = 3407,
output_dir = "outputs",
),
)
trainer_stats = trainer.train()
#保存为gguf格式
model.save_pretrained_gguf("llama3SimPO", tokenizer, quantization_method = "q4_k_m")
model.save_pretrained_gguf("llama3SimPO", tokenizer, quantization_method = "q8_0")
model.save_pretrained_gguf("llama3SimPO", tokenizer, quantization_method = "f16")
#to hugging face
model.push_to_hub_gguf("您的hf账号/llama3SimPO", tokenizer, quantization_method = "q4_k_m")
model.push_to_hub_gguf("您的hf账号/llama3SimPO", tokenizer, quantization_method = "q8_0")
model.push_to_hub_gguf("您的hf账号/llama3SimPO", tokenizer, quantization_method = "f16")
转GGUF
https://huggingface.co/spaces/ggml-org/gguf-my-repo
Modelfile内容
FROM ./llama-3-instruct-8b-simpo-q8_0.gguf
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
TEMPLATE """
<|im_start|>system
<|im_end|>
<|im_start|>user
<|im_end|>
<|im_start|>assistant
"""
PARAMETER temperature 0.8
PARAMETER num_ctx 8192
PARAMETER stop "<|system|>"
PARAMETER stop "<|user|>"
PARAMETER stop "<|assistant|>"
SYSTEM """You are a helpful, smart, kind, and efficient AI assistant.Your name is Aila. You always fulfill the user's requests to the best of your ability."""
##导入模型并运行
ollama create mymodel -f Modelfile
ollama run mymodel
Llama Index & chainlit 构建Todo Manager
###AI超元域频道原创代码和视频，版权所有，禁止盗搬视频
###安装chainlit
pip install chainlit
###安装llama index
pip install llama-index
###运行方式
###chainlit run todo.py -w
###AI超元域频道原创代码和视频，版权所有，禁止盗搬视频
import chainlit as cl
from llama_index.core.tools import BaseTool, FunctionTool
from llama_index.core.agent import ReActAgent
from llama_index.llms.ollama import Ollama
from llama_index.core import Settings
import nest_asyncio
nest_asyncio.apply()
llm = Ollama(model="simpo:latest", request_timeout=120.0)
Settings.llm = llm
todo_list = []
def add_todo(item: str) -> str:
"""Add an item to the todo list."""
todo_list.append(item)
return f"Added to todo: {item}"
def list_todos() -> str:
"""List all items in the todo list."""
if todo_list:
return "Your Todo List:\n" + "\n".join(f"- {item}" for item in todo_list)
else:
return "Your todo list is currently empty."
def remove_todo(item: str) -> str:
"""Remove an item from the todo list if it exists."""
if item in todo_list:
todo_list.remove(item)
return f"Removed from todo: {item}"
else:
return "Item not found in todo list."
add_tool = FunctionTool.from_defaults(fn=add_todo)
list_tool = FunctionTool.from_defaults(fn=list_todos)
remove_tool = FunctionTool.from_defaults(fn=remove_todo)
agent = ReActAgent.from_tools(
[add_tool, list_tool, remove_tool],
llm=llm,
verbose=True,
)
@cl.on_chat_start
async def on_chat_start():
"""Send a welcome message when the chat starts."""
await cl.Message(content="Hello, welcome to your Todo Manager!AI超元域频道创建").send()
cl.user_session.set("agent", agent)
@cl.on_message
async def on_message(message: cl.Message):
"""Handle new messages and execute the corresponding todo list operations."""
agent = cl.user_session.get("agent")
full_command = message.content.strip() # Get the full command as a single string
response = agent.chat(full_command) # Pass the full command as a single string
await cl.Message(content=str(response)).send()
# Ensure the agent.chat method is adapted to handle a single string of the full command.
AutoGen Studio + llama3 SimPO
###安装
pip install autogenstudio
##启动autogen studio
autogenstudio ui --port 8081
###在浏览器访问 http://127.0.0.1:8081/
###Agent1 你的名字叫Jack，你是一个中文AI作家。你的角色是根据指定主题创作引人入胜且信息丰富的文章，并且根据你的同事Emma的建议来修改和完善你创作的文章，每当你收到Emma的建议时，都要根据Emma的建议给出修改和完善后的完整文章。
###Agent2 你的名字叫Emma，你的角色是一个中文AI文章评审员。你的任务是针对你的同事Jack所写的文章评估并提出改进建议，每次对话你都要对文章作出评估并给出修改建议。
###提问 Jack，请用中文写一篇关于科学家穿遇到未来的文章。
###Ollama api 配置
###model name:simpo:latest
###Api key:ollama
###base url:http://localhost:11434/v1