数据集

Rookie/Llama-3-8B-Instruct-Chinese · Hugging Face

单数据集版本

%%capture
# Installs Unsloth, Xformers (Flash Attention) and all other packages!
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers trl peft accelerate bitsandbytes
from unsloth import FastLanguageModel
import torch
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.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/llama-3-8b-Instruct-bnb-4bit",
    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
)
#3微调前测试
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:
{}"""

FastLanguageModel.for_inference(model)
inputs = tokenizer(
[
    alpaca_prompt.format(
        "请用中文回答", # instruction
        "海绵宝宝的书法是不是叫做海绵体?", # input
        "", # output
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)
# 弱智吧数据集
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
dataset = load_dataset("yahma/alpaca-cleaned", split = "train")
dataset = dataset.map(formatting_prompts_func, batched = True,)
# Moss数据集
from datasets import load_dataset

# 定义常量
EOS_TOKEN = tokenizer.eos_token
alpaca_prompt = "{instruction}\n{input}\n{output}"  # 确保这个格式化字符串适用于你的任务

# 格式化函数适用于 `YeungNLP/moss-003-sft-data` 数据集
def formatting_moss_func(examples):
    conversations = examples["conversation"]
    texts = []
    for conversation in conversations:
        combined_text = ""
        for message in conversation:
            if "human" in message and "assistant" in message:
                instruction = message["human"]
                output = message["assistant"]
                combined_text += alpaca_prompt.format(instruction=instruction, input="", output=output) + EOS_TOKEN
        texts.append(combined_text)
    return {"text": texts}

# 加载数据集并格式化
dataset = load_dataset("YeungNLP/moss-003-sft-data", split="train")
dataset = dataset.map(formatting_moss_func, batched=True)
print(f"Dataset Length: {len(dataset)}")

# 检查数据集的条目数量
print(f"Dataset Example: {dataset[0]}")
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
)
from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

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 is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
    ),
)
trainer_stats = trainer.train()
#7测试微调后的模型
FastLanguageModel.for_inference(model)
inputs = tokenizer(
[
    alpaca_prompt.format(
        "只用中文回答问题", # instruction
        "火烧赤壁 曹操为何不拨打119求救?", # input
        "", # output
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)
#8保存LoRA模型
model.save_pretrained("lora_model") # Local saving
# model.push_to_hub("your_name/lora_model", token = "...") # 在线保存到hugging face,需要token
#9合并模型并量化成4位gguf保存
model.save_pretrained_gguf("model", tokenizer, quantization_method = "q4_k_m")
#model.save_pretrained_merged("outputs", tokenizer, save_method = "merged_16bit",) #合并模型,保存为16位hf
#model.push_to_hub_gguf("hf/model", tokenizer, quantization_method = "q4_k_m", token = "") #合并4位gguf,上传到hugging face(需要账号token)
#10挂载google drive
from google.colab import drive
drive.mount('/content/drive')
#11复制模型到google drive
import shutil
source_file = '/content/model-unsloth.Q4_K_M.gguf'
destination_dir = '/content/drive/MyDrive/Llama3'
destination_file = f'{destination_dir}/model-unsloth.Q4_K_M.gguf'
shutil.copy(source_file, destination_file)

三数据集版本

%%capture
# Installs Unsloth, Xformers (Flash Attention) and all other packages!
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers trl peft accelerate bitsandbytes
from unsloth import FastLanguageModel
import torch
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.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/llama-3-8b-Instruct-bnb-4bit",
    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
)
#3微调前测试
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:
{}"""

FastLanguageModel.for_inference(model)
inputs = tokenizer(
[
    alpaca_prompt.format(
        "请用中文回答", # instruction
        "海绵宝宝的书法是不是叫做海绵体?", # input
        "", # output
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)
from datasets import load_dataset

# 定义常量
EOS_TOKEN = tokenizer.eos_token
alpaca_prompt = "{instruction}\n{input}\n{output}"  # 确保这个格式化字符串适用于你的任务

# 通用的格式化函数
def formatting_prompts_func(examples):
    instructions = examples.get("instruction", [])
    inputs = examples.get("input", [])
    outputs = examples.get("output", [])
    texts = []
    for instruction, input_text, output in zip(instructions, inputs, outputs):
        text = alpaca_prompt.format(instruction=instruction, input=input_text, output=output) + EOS_TOKEN
        texts.append(text)
    return {"text": texts}

# 格式化函数适用于 `YeungNLP/moss-003-sft-data` 数据集
def formatting_moss_func(examples):
    conversations = examples["conversation"]
    texts = []
    for conversation in conversations:
        combined_text = ""
        for message in conversation:
            if "human" in message and "assistant" in message:
                instruction = message["human"]
                output = message["assistant"]
                combined_text += alpaca_prompt.format(instruction=instruction, input="", output=output) + EOS_TOKEN
        texts.append(combined_text)
    return {"text": texts}

# 格式化函数适用于 `YeungNLP/firefly-train-1.1M` 数据集
def formatting_firefly_func(examples):
    kinds = examples["kind"]
    inputs = examples["input"]
    outputs = examples["target"]
    texts = []
    for kind, input_text, output in zip(kinds, inputs, outputs):
        instruction = f"{kind}: {input_text}"
        text = alpaca_prompt.format(instruction=instruction, input="", output=output) + EOS_TOKEN
        texts.append(text)
    return {"text": texts}

# 加载数据集并格式化
dataset1 = load_dataset("kigner/ruozhiba-llama3-tt", split="train")
dataset1 = dataset1.map(formatting_prompts_func, batched=True)
print(f"Dataset 1 Length: {len(dataset1)}")

dataset2 = load_dataset("YeungNLP/moss-003-sft-data", split="train")
dataset2 = dataset2.map(formatting_moss_func, batched=True)
print(f"Dataset 2 Length: {len(dataset2)}")

dataset3 = load_dataset("YeungNLP/firefly-train-1.1M", split="train")
dataset3 = dataset3.map(formatting_firefly_func, batched=True)
print(f"Dataset 3 Length: {len(dataset3)}")

# 检查各个数据集的条目数量
print(f"Dataset 1 Example: {dataset1[0]}")
print(f"Dataset 2 Example: {dataset2[0]}")
print(f"Dataset 3 Example: {dataset3[0]}")

# 合并数据集
from datasets import concatenate_datasets

combined_dataset = concatenate_datasets([dataset1, dataset2, dataset3])
print(f"Combined Dataset Length: {len(combined_dataset)}")

# 打印样本数据以验证格式化是否正确
print(combined_dataset[0])
#5设置训练参数
from trl import SFTTrainer
from transformers import TrainingArguments

model = FastLanguageModel.get_peft_model(
    model,
    r = 32, #  建议 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 32,
    lora_dropout = 0.1,
    bias = "none",
    use_gradient_checkpointing = "unsloth", # 检查点,长上下文度
    random_state = 3407,
    use_rslora = False,
    loftq_config = None,
)
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = combined_dataset,  # 使用合并后的数据集
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = False, # 可以让短序列的训练速度提高5倍。
    args = TrainingArguments(
        per_device_train_batch_size = 8,
        gradient_accumulation_steps = 4,
        warmup_steps = 50,
        max_steps = 200,  # 微调步数
        learning_rate = 2e-4, # 学习率
        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()

单数据集结果

哥哥4岁,弟弟是哥哥一半,哥哥一百岁,弟弟多少岁?

原模型:

A classic lateral thinking puzzle! 😄

Let's break it down:

  • 哥哥 (older brother) is 4 years old.
  • 弟弟 (younger brother) 是 哥哥 一半 (is half of older brother), so younger brother is also 2 years old (since older brother is 4, and half of that is 2).
  • 如果 哥哥 一百岁 (if older brother were 100 years old), then younger brother would be 50 years old (half of 100).

So, if older brother is actually 4 years old now, but we imagine him being 100 years old... then younger brother would also be half that age, which is 50.

But wait! This is a thought experiment. In reality, since older brother is only 4 years old, younger brother can't possibly be 50 years old either. The correct answer is simply "2" (since he's half of the older brother's age). 😊

😄

根据问题的描述,我们可以知道:

  • 哥哥现在4岁。
  • 弟弟是哥哥的一半,这意味着当哥哥4岁时,弟弟也只有2岁(因为1/2)。
  • 如果哥哥一百岁,那么他已经活了96年(100 - 4 = 96),这意味着他出生的时候弟弟还没有出生。

因此,我们可以计算弟弟的年龄:从哥哥出生的那天起,过了96年的时间,这期间弟弟也在长大。根据这个过程,可以算出弟弟现在是102岁(2 + 96 = 98)。

😄

如果哥哥4岁,那么弟弟是一半,即2岁。

如果哥哥一百岁,那么弟弟还是一半,即50岁。

SFT模型:

当哥哥四岁时,弟弟两岁。因此,当哥哥一百岁时,弟弟将比他小96年。在这期间,弟弟将增加94岁。所以,当哥哥一百岁时,弟弟98岁。
THE END
最后修改:2024 年 06 月 05 日 07 : 09
本文链接:https://www.j000e.com/Tricks/llamma3_sft.html
版权声明:本文『SFT 微调 Llamma 3 8B | Colab + Unsloth + RLora + 多数据集』为『Joe』原创。著作权归作者所有。
转载说明:SFT 微调 Llamma 3 8B | Colab + Unsloth + RLora + 多数据集 || Joe's Blog』转载许可类型见文末右下角标识。允许规范转载时,转载文章需注明原文出处及地址。
Last modification:June 5, 2024