Submitting a Fine-Tuning Job: Organising the Workforce-拾光赋

Letters, Legacy, and Learning: Fine-Tuning LLMs Inspired by the Six Triple Eight (8 Part Series)

1 Six Triple Eight Redux: Fine-Tuning LLMs to Tackle Impossible Mail Mysteries of WWII
2 Exploratory Data Analysis: Digging Through the Backlog
… 4 more parts…
3 Counting Tokens: Sorting Through the Details
4 Data Splitting: Breaking Down the Problem
5 Understanding the OpenAI JSONL Format: Organising the Records
6 Uploading Files to OpenAI: Passing the Baton
7 Submitting a Fine-Tuning Job: Organising the Workforce
8 Inference with Fine-Tuned Models: Delivering the Message

The Six Triple Eight relied on discipline and coordination to execute their mission. We’ll mirror this by creating and submitting a fine-tuning job, allowing the LLM to learn from our curated dataset.

Fine-Tuning with OpenAI

When you create a fine-tuning job via client.fine_tuning.job.create(), you submit your configuration and dataset to OpenAI for training. Below are the key parameters and their purposes.

1. Parameters Overview

model

Description: The pre-trained GPT model you wish to fine-tune.
Examples: "gpt-3.5-turbo", "davinci", "gpt-4-mini" (hypothetical).

training_file

Description: The file ID of an uploaded JSONL file containing your training data.
Note: Obtain this ID by uploading your dataset with the Files API and storing the file_id.

hyperparameters

Description: A dictionary specifying the fine-tuning hyperparameters.
Key Fields:
- batch_size: Number of examples per batch (auto by default).
- learning_rate_multiplier: Scale factor for the learning rate (auto by default).
- n_epochs: Number of epochs (passes through the entire dataset).

suffix

Description: A custom string (up to 18 characters) appended to the fine-tuned model name.

seed

Description: Integer for reproducibility.
Usage: Ensures the same randomization and consistent training results across runs.

validation_file

Description: The file ID of a JSONL file containing your validation set.
Optional: But recommended for tracking overfitting and ensuring a well-generalized model.

integrations

Description: A list of integrations (e.g., Weights & Biases) you want enabled for the job.
Fields: Typically includes type and integration-specific configurations.


<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>job</span><span>.</span><span>create</span><span>(</span>
    <span>model</span><span>=</span><span>"</span><span>gpt-3.5-turbo</span><span>"</span><span>,</span>
    <span>training_file</span><span>=</span><span>"</span><span>train_id</span><span>"</span><span>,</span>
    <span>hyperparameters</span><span>=</span><span>{</span>
        <span>"</span><span>n_epochs</span><span>"</span><span>:</span> <span>1</span>
    <span>},</span>
    <span>validation_file</span><span>=</span><span>"</span><span>val_id</span><span>"</span>
<span>)</span>
<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>job</span><span>.</span><span>create</span><span>(</span>
    <span>model</span><span>=</span><span>"</span><span>gpt-3.5-turbo</span><span>"</span><span>,</span>
    <span>training_file</span><span>=</span><span>"</span><span>train_id</span><span>"</span><span>,</span>
    <span>hyperparameters</span><span>=</span><span>{</span>
        <span>"</span><span>n_epochs</span><span>"</span><span>:</span> <span>1</span>
    <span>},</span>
    <span>validation_file</span><span>=</span><span>"</span><span>val_id</span><span>"</span>
<span>)</span>
client.fine_tuning.job.create(
    model="gpt-3.5-turbo",
    training_file="train_id",
    hyperparameters={
        "n_epochs": 1
    },
    validation_file="val_id"
)

Enter fullscreen mode Exit fullscreen mode

Managing Fine-Tuning Jobs
Retrieves up to 10 fine-tuning jobs.


<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>jobs</span><span>.</span><span>list</span><span>(</span><span>limit</span><span>=</span><span>10</span><span>)</span>
<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>jobs</span><span>.</span><span>list</span><span>(</span><span>limit</span><span>=</span><span>10</span><span>)</span>
client.fine_tuning.jobs.list(limit=10)

Enter fullscreen mode Exit fullscreen mode

Retrieve a Specific Job


<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>retrieve</span><span>(</span><span>"</span><span>job_id</span><span>"</span><span>)</span>
<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>retrieve</span><span>(</span><span>"</span><span>job_id</span><span>"</span><span>)</span>
client.fine_tuning.retrieve("job_id")

Enter fullscreen mode Exit fullscreen mode

List Events for a Job


<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>list_events</span><span>(</span>
    <span>fine_tuning_job_id</span><span>=</span><span>"</span><span>xxxx</span><span>"</span><span>,</span>
    <span>limit</span><span>=</span><span>5</span>
<span>)</span>
<span>client</span><span>.</span><span>fine_tuning</span><span>.</span><span>list_events</span><span>(</span>
    <span>fine_tuning_job_id</span><span>=</span><span>"</span><span>xxxx</span><span>"</span><span>,</span>
    <span>limit</span><span>=</span><span>5</span>
<span>)</span>
client.fine_tuning.list_events(
    fine_tuning_job_id="xxxx",
    limit=5
)

Enter fullscreen mode Exit fullscreen mode

Summary

Model Selection: Choose a suitable GPT model to fine-tune.
Data Preparation: Upload JSONL files and note their IDs.
Hyperparameters: Tune batch size, learning rate, and epochs for optimal performance.
Monitoring: Use validation files, job retrieval, and event logging to ensure your model trains effectively.
Reproducibility: Set a seed if consistent results are important for your workflow.
By following these steps, you’ll have a clear path to submitting and managing your fine-tuning jobs in OpenAI, ensuring your model is trained precisely on your custom data.

Letters, Legacy, and Learning: Fine-Tuning LLMs Inspired by the Six Triple Eight (8 Part Series)

原文链接：Submitting a Fine-Tuning Job: Organising the Workforce

展开阅读全文

文章版权声明 1、本网站名称：拾光赋
2、本站永久网址：https://www.blogs.ink
3、本网站的文章部分内容可能来源于网络，仅供大家学习与参考，如有侵权，请联系站长QQ：805375623进行删除处理。
4、本站一切资源不代表本站立场，并不代表本站赞同其观点和对其真实性负责。
5、本站一律禁止以任何方式发布或转载任何违法的相关信息，访客发现请向站长举报
6、本站资源大多存储在云盘，如发现链接失效，请联系我们我们会第一时间更新。

THE END