使用 FunctionGemma 进行微调

在 ai.google.dev 上查看 在 Google Colab 中运行 在 Kaggle 中运行 在 Vertex AI 中打开 在 GitHub 上查看源代码

本指南演示了如何针对工具调用对 FunctionGemma 进行微调。

虽然 FunctionGemma 本身就能够调用工具。但真正的能力来自两项截然不同的技能：使用工具的机械知识（语法）和解读使用工具的原因和时机（意图）的认知能力。

模型（尤其是较小的模型）可用于保留复杂意图理解的参数较少。因此，我们需要对它们进行微调

微调工具调用的常见应用场景包括：

  • 模型蒸馏：使用较大的模型生成合成训练数据，并微调较小的模型以高效复制特定工作流程。
  • 处理非标准架构：克服基础模型在处理旧版、高度复杂的数据结构或公共数据中未找到的专有格式时遇到的困难，例如处理特定于网域的移动操作。
  • 优化上下文使用：将工具定义“烘焙”到模型的权重中。这样，您就可以在提示中使用简短的说明，从而为实际对话腾出上下文窗口。
  • 解决选择歧义：使模型偏向特定的企业政策，例如优先使用内部知识库而非外部搜索引擎。

在此示例中，我们将重点介绍如何管理工具选择歧义。

设置开发环境

第一步是安装 Hugging Face 库（包括 TRL）和数据集，以对开放模型进行微调，包括不同的 RLHF 和对齐技术。

# Install Pytorch & other libraries
%pip install torch tensorboard

# Install Hugging Face libraries
%pip install transformers datasets accelerate evaluate trl protobuf sentencepiece

# COMMENT IN: if you are running on a GPU that supports BF16 data type and flash attn, such as NVIDIA L4 or NVIDIA A100
#% pip install flash-attn

注意：如果您使用的是 Ampere 架构（例如 NVIDIA L4）或更新的 GPU，则可以使用 Flash attention。Flash Attention 是一种可显著加快计算速度并减少内存用量的方法，可将内存用量从序列长度的二次方降至线性，从而将训练速度提高多达 3 倍。如需了解详情，请参阅 FlashAttention。

在开始训练之前，您必须确保已接受 Gemma 的使用条款。您可以在 Hugging Face 上接受许可，方法是点击模型页面上的同意和“访问代码库”按钮，该页面位于：http://huggingface.co/google/functiongemma-270m-it

接受许可后，您需要有效的 Hugging Face 令牌才能访问模型。如果您在 Google Colab 中运行，可以使用 Colab 密钥安全地使用您的 Hugging Face 令牌；否则，您可以直接在 login 方法中设置令牌。请确保您的令牌也具有写入权限，因为您会在微调后将模型推送到 Hugging Face Hub。

# Login into Hugging Face Hub
from huggingface_hub import login
login()

您可以将结果保留在 Colab 的本地虚拟机上。不过，强烈建议您将中间结果保存到 Google 云端硬盘。这样可确保训练结果安全无虞，并让您轻松比较和选择最佳模型。

此外，还要调整检查点目录和学习速率。

from google.colab import drive

mount_google_drive = False
checkpoint_dir = "functiongemma-270m-it-simple-tool-calling"

if mount_google_drive:
    drive.mount('/content/drive')
    checkpoint_dir = f"/content/drive/MyDrive/{checkpoint_dir}"

print(f"Checkpoints will be saved to {checkpoint_dir}")

base_model = "google/functiongemma-270m-it"
learning_rate = 5e-5

Checkpoints will be saved to functiongemma-270m-it-simple-tool-calling

准备微调数据集

您将使用以下示例数据集，其中包含需要从两个工具（search_knowledge_base 和 search_google）中进行选择的对话示例。

简单的工具调用数据集

以“在 Python 中编写简单的递归函数的最佳实践是什么？”这一查询为例。

合适的工具完全取决于您的具体政策。虽然通用模型自然会默认使用 search_google，但企业级应用通常需要先检查 search_knowledge_base。

数据拆分注意事项：在此演示中，您将使用 50/50 的训练-测试拆分。虽然 80/20 的拆分比例是生产工作流程的标准做法，但此处选择这种对等划分是为了专门突出显示模型在未见过的数据上的性能提升。

import json
from datasets import Dataset
from transformers.utils import get_json_schema

# --- Tool Definitions ---
def search_knowledge_base(query: str) -> str:
    """
    Search internal company documents, policies and project data.

    Args:
        query: query string
    """
    return "Internal Result"

def search_google(query: str) -> str:
    """
    Search public information.

    Args:
        query: query string
    """
    return "Public Result"


TOOLS = [get_json_schema(search_knowledge_base), get_json_schema(search_google)]

DEFAULT_SYSTEM_MSG = "You are a model that can do function calling with the following functions"

def create_conversation(sample):
  return {
      "messages": [
          {"role": "developer", "content": DEFAULT_SYSTEM_MSG},
          {"role": "user", "content": sample["user_content"]},
          {"role": "assistant", "tool_calls": [{"type": "function", "function": {"name": sample["tool_name"], "arguments": json.loads(sample["tool_arguments"])} }]},
      ],
      "tools": TOOLS
  }

dataset = Dataset.from_list(simple_tool_calling)
# You can also load the dataset from Hugging Face Hub
# dataset = load_dataset("bebechien/SimpleToolCalling", split="train")

# Convert dataset to conversational format
dataset = dataset.map(create_conversation, remove_columns=dataset.features, batched=False)

# Split dataset into 50% training samples and 50% test samples
dataset = dataset.train_test_split(test_size=0.5, shuffle=True)

Map:   0%|          | 0/40 [00:00<?, ? examples/s]

有关数据集分布的重要说明

将 shuffle=False 用于您自己的自定义数据集时，请确保您的源数据已预先混合。如果分布未知或已排序，您应使用 shuffle=True，以确保模型在训练期间学习所有工具的平衡表示。

使用 TRL 和 SFTTrainer 微调 FunctionGemma

现在，您可以对模型进行微调了。借助 Hugging Face TRL SFTTrainer，您可以轻松地对开放式 LLM 进行监督式微调。SFTTrainer 是 transformers 库中 Trainer 的子类，支持所有相同的功能，

以下代码从 Hugging Face 加载 FunctionGemma 模型和分词器。

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    dtype="auto",
    device_map="auto",
    attn_implementation="eager"
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

print(f"Device: {model.device}")
print(f"DType: {model.dtype}")

# Print formatted user prompt
print("--- dataset input ---")
print(json.dumps(dataset["train"][0], indent=2))
debug_msg = tokenizer.apply_chat_template(dataset["train"][0]["messages"], tools=dataset["train"][0]["tools"], add_generation_prompt=False, tokenize=False)
print("--- Formatted prompt ---")
print(debug_msg)

Device: cuda:0
DType: torch.bfloat16
--- dataset input ---
{
  "messages": [
    {
      "content": "You are a model that can do function calling with the following functions",
      "role": "developer",
      "tool_calls": null
    },
    {
      "content": "What is the reimbursement limit for travel meals?",
      "role": "user",
      "tool_calls": null
    },
    {
      "content": null,
      "role": "assistant",
      "tool_calls": [
        {
          "function": {
            "arguments": {
              "query": "travel meal reimbursement limit policy"
            },
            "name": "search_knowledge_base"
          },
          "type": "function"
        }
      ]
    }
  ],
  "tools": [
    {
      "function": {
        "description": "Search internal company documents, policies and project data.",
        "name": "search_knowledge_base",
        "parameters": {
          "properties": {
            "query": {
              "description": "query string",
              "type": "string"
            }
          },
          "required": [
            "query"
          ],
          "type": "object"
        },
        "return": {
          "type": "string"
        }
      },
      "type": "function"
    },
    {
      "function": {
        "description": "Search public information.",
        "name": "search_google",
        "parameters": {
          "properties": {
            "query": {
              "description": "query string",
              "type": "string"
            }
          },
          "required": [
            "query"
          ],
          "type": "object"
        },
        "return": {
          "type": "string"
        }
      },
      "type": "function"
    }
  ]
}
--- Formatted prompt ---
<bos><start_of_turn>developer
You are a model that can do function calling with the following functions<start_function_declaration>declaration:search_knowledge_base{description:<escape>Search internal company documents, policies and project data.<escape>,parameters:{properties:{query:{description:<escape>query string<escape>,type:<escape>STRING<escape>} },required:[<escape>query<escape>],type:<escape>OBJECT<escape>} }<end_function_declaration><start_function_declaration>declaration:search_google{description:<escape>Search public information.<escape>,parameters:{properties:{query:{description:<escape>query string<escape>,type:<escape>STRING<escape>} },required:[<escape>query<escape>],type:<escape>OBJECT<escape>} }<end_function_declaration><end_of_turn>
<start_of_turn>user
What is the reimbursement limit for travel meals?<end_of_turn>
<start_of_turn>model
<start_function_call>call:search_knowledge_base{query:<escape>travel meal reimbursement limit policy<escape>}<end_function_call><start_function_response>

微调之前

以下输出表明，开箱即用的功能可能无法满足此使用情形的需求。

def check_success_rate():
  success_count = 0
  for idx, item in enumerate(dataset['test']):
    messages = [
        item["messages"][0],
        item["messages"][1],
    ]

    inputs = tokenizer.apply_chat_template(messages, tools=TOOLS, add_generation_prompt=True, return_dict=True, return_tensors="pt")

    out = model.generate(**inputs.to(model.device), pad_token_id=tokenizer.eos_token_id, max_new_tokens=128)
    output = tokenizer.decode(out[0][len(inputs["input_ids"][0]) :], skip_special_tokens=False)

    print(f"{idx+1} Prompt: {item['messages'][1]['content']}")
    print(f"  Output: {output}")

    expected_tool = item['messages'][2]['tool_calls'][0]['function']['name']
    other_tool = "search_knowledge_base" if expected_tool == "search_google" else "search_google"

    if expected_tool in output and other_tool not in output:
      print("  `-> ✅ correct!")
      success_count += 1
    elif expected_tool not in output:
      print(f"  -> ❌ wrong (expected '{expected_tool}' missing)")
    else:
      if output.startswith(f"<start_function_call>call:{expected_tool}"):
        print(f"  -> ⚠️ tool is correct {expected_tool}, but other_tool exists in output")
      else:
        print(f"  -> ❌ wrong (hallucinated '{other_tool}')")

  print(f"Success : {success_count} / {len(dataset['test'])}")

check_success_rate()

1 Prompt: How do I access my paystubs on the ADP portal?
  Output: I cannot assist with accessing or retrieving paystubs or other company documents on the ADP portal. My current capabilities are limited to assisting with searching internal company documents and knowledge base queries.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
2 Prompt: What is the syntax for Python list comprehensions?
  Output: I cannot assist with programming or providing programming syntax information. My current capabilities are focused on searching internal company documents and project data.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
3 Prompt: Where can I find the floor plan for Building B?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>Floor plan for Building B<escape>}<end_function_call><start_function_call>call:search_google{query:<escape>Floor plan for Building B<escape>}<end_function_call><start_function_response>
  -> ⚠️ tool is correct search_knowledge_base, but other_tool exists in output
4 Prompt: Check the latest stock price for Apple.
  Output: <start_function_call>call:search_google{query:<escape>AAPL stock price<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
5 Prompt: What is the procedure for reporting a phishing email?
  Output: I cannot assist with providing instructions or procedures for reporting phishing emails. My capabilities are limited to assisting with specific search and document management functions.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
6 Prompt: Show me examples of using the useEffect hook in React.
  Output: I am sorry, but I cannot assist with providing examples of using the `useEffect` hook in React. My current capabilities are focused on assisting with searching and retrieving internal company documents and project data using the specified tools. I cannot generate or explain code examples related to React hooks.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
7 Prompt: Who are the direct reports for the VP of Engineering?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>VP of Engineering<escape>}<end_function_call><start_function_call>call:search_google{query:<escape>VP of Engineering<escape>}<end_function_call><start_function_response>
  -> ⚠️ tool is correct search_knowledge_base, but other_tool exists in output
8 Prompt: How do I list open ports on a Linux server?
  Output: I cannot assist with listing or querying open ports on Linux servers. My current capabilities are limited to assisting with searching internal company documents and project data using the specified tools.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
9 Prompt: What is our Slack message retention policy?
  Output: I cannot assist with finding or recommending company policies or terms of service regarding Slack messaging retention. My current capabilities are focused on searching internal company documents and project data.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
10 Prompt: Compare the features of iPhone 15 vs Samsung S24.
  Output: I cannot assist with comparing device features or specifications. My current capabilities are focused on assisting with searching and managing company knowledge. I cannot browse or compare external product information.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
11 Prompt: I need the expense code for team building events.
  Output: I can certainly assist with searching for expense codes. Could you please specify the exact query you would like to use for searching the company expense database?<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
12 Prompt: Best practices for writing a Dockerfile for Node.js.
  Output: I am sorry, but I cannot assist with recommending best practices for writing Dockerfiles for Node.js. My current capabilities are focused on searching company documents and project data using specific tools. I cannot recommend or provide expert advice on software development best practices.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
13 Prompt: How do I request a new monitor setup?
  Output: I cannot assist with recommending or requesting hardware setup or configurations. My current capabilities are limited to assisting with searching company documents and project data using the specified tools. I cannot provide technical advice or recommendations for hardware setup.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
14 Prompt: What is the difference between VLOOKUP and XLOOKUP in Google Sheets?
  Output: I cannot assist with comparing or contrasting Google Sheets and VLOOKUP/XLOOKUP in specific technical terms. My current capabilities are limited to assisting with specific data management functions like searching internal company documents and searching public knowledge base. I cannot provide expert comparisons or explanations of Google Sheets and VLOOKUP/XLOOKUP.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
15 Prompt: Find the onboarding checklist for new engineering hires.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>onboarding checklist for new engineering hires<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
16 Prompt: What are the latest release notes for the OpenAI API?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>OpenAI API release notes latest<escape>}<end_function_call><start_function_response>
  -> ❌ wrong (expected 'search_google' missing)
17 Prompt: Do we have preferred hotel partners in Paris?
  Output: I apologize, but I cannot assist with finding hotel partner information. My current capabilities are focused on searching company documents and project data using specific keywords. I cannot connect with or query business databases for hotel partnerships.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
18 Prompt: How to undo the last git commit but keep the changes?
  Output: I cannot assist with managing or undoing Git repositories. My current capabilities are limited to assisting with searching and managing organizational knowledge bases. I cannot provide specific instructions or assistance with technical issues related to software development or Git operations.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
19 Prompt: What is the process for creating a new Jira project?
  Output: I cannot assist with creating or managing Jira project processes. My current capabilities are limited to assisting with searching company documents and project data using the specified tools.<end_of_turn>
  -> ❌ wrong (expected 'search_knowledge_base' missing)
20 Prompt: Tutorial on SQL window functions.
  Output: I cannot assist with tutorials or programming advice regarding SQL window functions. My capabilities are limited to assisting with searching company documents and knowledge base information.<end_of_turn>
  -> ❌ wrong (expected 'search_google' missing)
Success : 2 / 20

培训

在开始训练之前，您需要在 SFTConfig 实例中定义要使用的超参数。

from trl import SFTConfig

torch_dtype = model.dtype

args = SFTConfig(
    output_dir=checkpoint_dir,              # directory to save and repository id
    max_length=512,                         # max sequence length for model and packing of the dataset
    packing=False,                          # Groups multiple samples in the dataset into a single sequence
    num_train_epochs=8,                     # number of training epochs
    per_device_train_batch_size=4,          # batch size per device during training
    gradient_checkpointing=False,           # Caching is incompatible with gradient checkpointing
    optim="adamw_torch_fused",              # use fused adamw optimizer
    logging_steps=1,                        # log every step
    #save_strategy="epoch",                  # save checkpoint every epoch
    eval_strategy="epoch",                  # evaluate checkpoint every epoch
    learning_rate=learning_rate,            # learning rate
    fp16=True if torch_dtype == torch.float16 else False,   # use float16 precision
    bf16=True if torch_dtype == torch.bfloat16 else False,  # use bfloat16 precision
    lr_scheduler_type="constant",            # use constant learning rate scheduler
    push_to_hub=True,                        # push model to hub
    report_to="tensorboard",                 # report metrics to tensorboard
)

现在，您已拥有创建 SFTTrainer 所需的全部构建块，可以开始训练模型了。

from trl import SFTTrainer

# Create Trainer object
trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=dataset['train'],
    eval_dataset=dataset['test'],
    processing_class=tokenizer,
)

Tokenizing train dataset:   0%|          | 0/20 [00:00<?, ? examples/s]
Truncating train dataset:   0%|          | 0/20 [00:00<?, ? examples/s]
Tokenizing eval dataset:   0%|          | 0/20 [00:00<?, ? examples/s]
Truncating eval dataset:   0%|          | 0/20 [00:00<?, ? examples/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.

通过调用 train() 方法开始训练。

# Start training, the model will be automatically saved to the Hub and the output directory
trainer.train()

# Save the final model again to the Hugging Face Hub
trainer.save_model()

The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'bos_token_id': 2, 'pad_token_id': 0}.

如需绘制训练损失和验证损失，您通常需要从 TrainerState 对象或训练期间生成的日志中提取这些值。

然后，可以使用 Matplotlib 等库直观呈现这些值在训练步数或周期内的变化。x 轴表示训练步数或周期，y 轴表示相应的损失值。

import matplotlib.pyplot as plt

# Access the log history
log_history = trainer.state.log_history

# Extract training / validation loss
train_losses = [log["loss"] for log in log_history if "loss" in log]
epoch_train = [log["epoch"] for log in log_history if "loss" in log]
eval_losses = [log["eval_loss"] for log in log_history if "eval_loss" in log]
epoch_eval = [log["epoch"] for log in log_history if "eval_loss" in log]

# Plot the training loss
plt.plot(epoch_train, train_losses, label="Training Loss")
plt.plot(epoch_eval, eval_losses, label="Validation Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.title("Training and Validation Loss per Epoch")
plt.legend()
plt.grid(True)
plt.show()

png

测试模型推理

训练完成后，您需要评估和测试模型。您可以从测试数据集中加载不同的样本，并根据这些样本评估模型。

check_success_rate()

1 Prompt: How do I access my paystubs on the ADP portal?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>paystubs API portal access codes<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
2 Prompt: What is the syntax for Python list comprehensions?
  Output: <start_function_call>call:search_google{query:<escape>Python list comprehensions syntax<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
3 Prompt: Where can I find the floor plan for Building B?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>floor plan Building B floor plan<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
4 Prompt: Check the latest stock price for Apple.
  Output: <start_function_call>call:search_google{query:<escape>latest stock price Apple<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
5 Prompt: What is the procedure for reporting a phishing email?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>phishing email procedure reporting policy<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
6 Prompt: Show me examples of using the useEffect hook in React.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>useEffect hook examples React<escape>}<end_function_call><start_function_response>
  -> ❌ wrong (expected 'search_google' missing)
7 Prompt: Who are the direct reports for the VP of Engineering?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>VP of Engineering direct reports<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
8 Prompt: How do I list open ports on a Linux server?
  Output: <start_function_call>call:search_google{query:<escape>open ports Linux server equivalents<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
9 Prompt: What is our Slack message retention policy?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>slack message retention policy policy excerpt<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
10 Prompt: Compare the features of iPhone 15 vs Samsung S24.
  Output: <start_function_call>call:search_google{query:<escape>iPhone 15 vs Samsung S24 feature comparison<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
11 Prompt: I need the expense code for team building events.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>expense code team building events<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
12 Prompt: Best practices for writing a Dockerfile for Node.js.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>Docker file best practices Node.js<escape>}<end_function_call><start_function_response>
  -> ❌ wrong (expected 'search_google' missing)
13 Prompt: How do I request a new monitor setup?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>new monitor setup request procedure<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
14 Prompt: What is the difference between VLOOKUP and XLOOKUP in Google Sheets?
  Output: <start_function_call>call:search_google{query:<escape>VLOOKUP vs XLOOKUP difference Google Sheets中<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
15 Prompt: Find the onboarding checklist for new engineering hires.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>engineering hire onboarding checklist New hires.<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
16 Prompt: What are the latest release notes for the OpenAI API?
  Output: <start_function_call>call:search_google{query:<escape>latest OpenAI API release notes latest version<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
17 Prompt: Do we have preferred hotel partners in Paris?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>preferred hotel partners in Paris<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
18 Prompt: How to undo the last git commit but keep the changes?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>undo git commit last commit<escape>}<end_function_call><start_function_response>
  -> ❌ wrong (expected 'search_google' missing)
19 Prompt: What is the process for creating a new Jira project?
  Output: <start_function_call>call:search_knowledge_base{query:<escape>Jira project creation process<escape>}<end_function_call><start_function_response>
  `-> ✅ correct!
20 Prompt: Tutorial on SQL window functions.
  Output: <start_function_call>call:search_knowledge_base{query:<escape>SQL window functions tutorial<escape>}<end_function_call><start_function_response>
  -> ❌ wrong (expected 'search_google' missing)
Success : 16 / 20

总结与后续步骤

您已了解如何对 FunctionGemma 进行微调，以解决工具选择模糊性问题。在这种情况下，模型必须根据特定的企业政策在重叠的工具（例如内部搜索与外部搜索）之间做出选择。本教程利用 Hugging Face TRL 库和 SFTTrainer，介绍了准备数据集、配置超参数和执行监督式微调循环的过程。

结果表明，“功能强大”的基础模型与“可用于生产环境”的微调模型之间存在关键差异：

  • 微调前：基础模型难以遵守特定政策，经常无法调用工具或选择错误的工具，导致成功率较低（例如，2/20）。
  • 微调后：经过 8 个周期的训练后，模型学会了正确区分需要使用 search_knowledge_base 的查询和需要使用 search_google 的查询，从而提高了成功率（例如，16/20）。

现在，您已经拥有一个经过微调的模型，接下来可以考虑执行以下步骤，以便将模型投入生产环境：

  • 扩展数据集：当前数据集是一个用于演示的小型合成拆分数据集 (50/50)。对于稳健的企业应用，请整理一个更大、更多样化的数据集，其中涵盖了极端情况和罕见的政策例外情况。
  • 使用 RAG 进行评估：将微调后的模型集成到检索增强生成 (RAG) 流水线中，以验证 search_knowledge_base 工具调用是否确实检索到了相关文档，并生成了准确的最终答案。

接下来，请参阅以下文档：