Guide: How to Fine-Tune GPT-3.5 for Translation

Fine-tuning GPT-3.5 for translation involves converting TMX/TBX files into JSONL, uploading training data to OpenAI, and launching a fine-tuning job via the UI or API. This process customizes terminology, improves accuracy, and brings performance closer to GPT-4o for translation tasks.

OpenAI platform offers its users an opportunity to fine-tune GPT-3.5. For localization needs, one can train the AI to generate translation output with specific terminology, translation memory, and style guides. After the training, the model will not be able to generate dialogue-like answers anymore and will become a tool only for translation.

When you train Gen AI for translation the impact is somewhat less predictable than it is with machine translation. However, enough fine-tuning iterations will eventually get GPT-3.5 to an output quality level comparable to GPT-4o, with improved terminology compliance. The whole training generally costs $10-$100 which is significantly less than training Google Translate or Microsoft Custom Translator. For example:

• OpenAI GPT-3.5 average fine-tune – $50
• Google Translate fine-tune – $45 / hour, average 6 hours, $300 max
• Microsoft Translator fine-tune – $10 per million source + target characters of training data (max. $300/training)

There are two ways to run the fine-tune, via Open AI UI and API calls. In the following guide, we will cover both options. Please note that it is possible to fine-tune GPT-4 and GPT-4o in the experimental access program. The workflow is the same, and eligible users can request access in the fine-tuning interface when creating a new fine-tuning job.

Creating the Data Set for AI Training

Unlike Google Translate and Microsoft Translator, which use glossaries in TBX and translation memories in TMX, OpenAI’s models need to be fine-tuned with JSON files. So the first step is to convert TMX and TBX into JSON. Here is the example JSON structure that you can use:

[
  {  // First JSON object
    "messages": [
      {
        "role": "system",
        "content": "system text 1"
      },

      {
        "role": "user",
        "content": "user text 1"
      },

      {
        "role": "assistant",
        "content": "assistant text 1"
      }
    ]
  },

  {  // Second JSON object
    "messages": [
      {
        "role": "system",
        "content": "system text 2"
      },

      {
        "role": "user",
        "content": "user text 2"
      },

      {
        "role": "assistant",
        "content": "assistant text 2"
      }
    ]
  },

  {  // Third JSON object (and so on...)
    "messages": [
      {
        "role": "system",
        "content": "system text 3"
      },

      {
        "role": "user",
        "content": "user text 3"
      },

      {
        "role": "assistant",
        "content": "assistant text 3"
      }
    ]
  }
]

At Custom.MT, we convert TMX into JSON with a custom Python script, which can be developed by a coder, or you can ask GPT-4o to write such a converter for you with a prompt that shows the example JSON structure above and a snippet of your translation memory. Here is what our converter looks like:

Only add high-quality sentences to OpenAI. Fine-tuning LLMs doesn’t require a lot of examples, and it is better to delete all data that you consider old, inconsistent, and risky. We use a proprietary tool for that, called Custom.MT TMX Cleaner, but there are multiple open options on the market, such as Heartsome, TMop, Trados Cleanup, etc.

Uploading a Training File to Open AI

Once your dataset is ready, the only thing to do is to upload it to OpenAI. There are two ways how to do it, via Open AI UI, and via the API.

To upload the dataset via the API, go to the OpenAPI account > Dashboard.  On the left menu, click Storage > Upload.

To upload programmatically, we recommend using a Python script. Programmatic uploads will enable you to launch fine-tuning jobs without manual operations, for example, if you’d like your TMS to automatically fine-tune regularly. Here is an example upload script in Python:

from openai import OpenAI

client = OpenAI(api_key="your-key")

client.files.create(
    file=open("en-fr-gpt.jsonl", "rb"),
    purpose="fine-tune"
)

Creating a Fine-Tuning Job

Again for this step, the controls are available via the user interface and the API. Here is how to create an OpenAI fine-tuning job via the UI:

From the left-hand menu, pick Fine-Tuning, then click Create on the right side of the screen. Chose your model from the pool of currently available for fine-tuning, and your JSON file.

JsonID can be found after clicking on the file you uploaded in the Storage,

Here is how to create a fine-tuning job in Python:

from openai import OpenAI

client = OpenAI()

client.fine_tuning.jobs.create(
  training_file="file-abc123",
  model="gpt-3.5-turbo"
)

Time to train the AI model depends on the size of the data set, typically we see 1-1.5 hour training time for fine-tuning jobs with smaller & cleaner translation memories with circa 1 million words in size.

Once the fine-tuning is complete, a green check mark will appear in the UI.

Translation Output after Training

Fine-tuned translation often shows greater accuracy and context alignment, using terminology that is more appropriate for the given context (e.g., “assistera à la conférence” vs. “participera à la conférence”). In the last example the translation remains unchanged since the first option complys with the requirements.

Do you want to use GenAI for translation at your company but are not quite sure where to start? Book a free consultation with our AI and MT translation experts to quickly gain a comprehensive understanding of how AI can best assist you: https://custom.mt/book-a-call/

Frequently Asked Questions