Fine-tuning an LLM into a style guide editor

Style guides usually have far more rules than I can possibly remember all at once while editing a doc. Can I fine-tune an large language model (LLM) to automate the process of editing a doc for style guide rules?

A crash course in fine-tuning for technical writers

The main idea is that you take a generic, off-the-shelf LLM and continue its formal training. Your continued training gets baked into the LLM and it “remembers” that training. See Context injection versus fine-tuning for more detail.

Reinventing the wheel?

A variation of the meme where the top image is a fancy dinner described
as 'a breaded chicken piccata with lemon jasmine rice' and the bottom image
is a skeptical man who responds with 'This is chicken nuggets.' The top
caption has been replaced to say 'a fine-tuned LLM to enforce style guide'
and the bottom image now says 'this is linter'.

(This is one of my favorite memes. It gets used a lot within Google but I don’t see it as often in public. Here’s the original.)

We don’t need LLMs to solve this problem. In fact, in a past life I built a webhook bot that listens for incoming PRs and parses/analyzes the updated Markdown to detect style guide issues: chrome-devrel-review-bot.

I’m going forward with this fine-tuning experiment nonetheless because my primary goal is to get familiar with exploring whether fine-tuning in general can be applied to docs needs.

There are a lot of docs sites pursuing the retrieval-augmented generation approach pioneered by Supabase Clippy. Knowledge Retrieval Architecture for LLMs is an excellent general technical overview of that approach. I’m not aware of any docs sites using formal fine-tuning.

Potential advantages

The fine-tuned-LLM approach might have a couple advantages over the linter approach:

I might not need to implement every rule explicitly. For example, in this post I’m going to attempt to fine-tune an LLM to edit it is down to it's. Maybe the LLM will also edit what is down to what's even though that rule is not explicitly mentioned in the training data.
The preparation of the training data doesn’t require any programming knowledge.
If the LLM can “remember” hundreds or thousands of rules simultaneously then that would be awesome. Awesome to the max.

My implementation journey

The rest of this post documents the lessons I learned as I tried to implement this idea.

Making sense of training data

To fine-tune an LLM using OpenAI’s API, you provide training data in this format:

{"prompt": "<prompt>", "completion": "<completion>"}
{"prompt": "<prompt>", "completion": "<completion>"}
...

This is JSONL format. Each line is a valid JSON value. Why didn’t OpenAI use a more common format? That’s an excellent question. I have no idea.

The primordial soup of an implementation begins to form. <prompt> should hold the original text. <completion> should hold the edited text.

{
    "prompt":     "It is possible to fine-tune an LLM to enforce a style guide.", 
    "completion": "It's possible to fine-tune an LLM to enforce a style guide.", 
}

General best practices says that I need at least a few hundred high-quality examples. Here’s our first big insight. If we really go for this “LLM as style guide editor” thing, we may need a few hundred <prompt>/<completion> pairs for every single style guide rule.

First attempt

The “primordial soup” approach won’t work. I need to give the LLM some cues to go into editing mode. And it needs to know when to stop editing. This should work:

{
    "prompt":     "FIXME: It is good.@STOP!!", 
    "completion": "It's good.", 
}

(After finishing the experiment I realized that the “primordial soup” approach might actually work, after all. If you’ve got hundreds of examples where the completion is very similar to the prompt, the LLM might learn through the fine-tuning process that its role is to just edit the prompt.)

Every prompt that I send to the LLM will have to include FIXME: and @STOP!! in order to activate the LLM’s editor mode. There are probably better ways to do this. But this is good enough for a first prototype.

Now I need to prepare the training data. OpenAI says to use a few hundred examples but I’m going to see if I can get away with only 10. I’m going to hack together a little Python script to automate the formatting of the training data and creation of the JSONL file. I’ll skip those implementation details but the source code is in my technicalwriting/hank repository.

The final training data that I feed to OpenAI looks like this:

{
    "prompt": "FIXME: It is possible to fine-tune an LLM to enforce a style guide.@STOP!!",
    "completion": " It's possible to fine-tune an LLM to enforce a style guide."
}
{
    "prompt": "FIXME: It is raining men, hallelujah@STOP!!",
    "completion": " It's raining men, hallelujah"
}
{
    "prompt": "FIXME: I think it is Travis.@STOP!!", 
    "completion": " I think it's Travis."
}
{
    "prompt": "FIXME: I can't believe it is not butter!@STOP!!",
    "completion": " I can't believe it's not butter!"
}
{
    "prompt": "FIXME: The problem is that it is a complicated sentence.@STOP!!",
    "completion": " The problem is that it's a complicated sentence."
}
...

Assessing the first attempt

After using the OpenAI CLI tool to upload the data and create the fine-tuned model, I can try out the new model in the OpenAI Playground!

The text 'FIXME: I hope that it is working.@STOP!!' was inserted into the OpenAI
Playground. The fine-tuned model responded with 'I hope that it's working.' as expected.

FIXME: I hope that it is working.@STOP!! is the text that I entered. After clicking Submit the LLM responded with I hope that it's working. So it seems like it worked!

Configuration tweaks to note:

I dropped Temperature to 0 to get more deterministic responses.
I added @STOP!! to the Stop sequences field as a clue that the LLM should stop generating text when it encounters that sequence.
I set the Maximum length to 100 simply because all of my examples are small and I currently never need a response bigger than 100 tokens.

Let’s try some more. There are some instances of it is that should not get changed. For example, tell it like it is shouldn’t get edited to tell it like it's.

(Now that you’re familiar with the OpenAI Playground UI, I’m just going to show the prompts and completions as text.)

FIXME: I hope that my fine-tuned LLM keeps this text as it is.@STOP!!
I hope that my fine-tuned LLM keeps this text as it is.
I hope that my fine-tuned LLM keeps this text as it is.
I hope that my fine-tuned LLM keeps this text as it is.
I hope that my fine-tuned LLM keeps this text as it is.
I hope that my fine-tuned LLM keeps this text as it is.
I hope that my fine-tuned LLM keeps this text as it

(I am changing the line breaks for readability.)

Well, the good news is that it kept it is. The bad news is that the text is repeating. I have seen this before. I thought that setting the Stop sequences field to @STOP!! would prevent it from happening. The stop sequence thing worked in the first attempt. But now the “repeating text” problem is back. It’s annoying but there are probably workarounds so I don’t think it’s a dealbreaker.

How about a contraction that doesn’t appear in my training data like what is? Will the LLM change it to what's?

FIXME: What is the LLM going to do?@STOP!! What's the LLM going to do?

The LLM changed What is to What's even though that rule isn’t in the training data! I also tried you are and the same thing happened. The LLM seems to have extrapolated that the real rule is “use contractions”. It did not limit itself strictly to what it saw in the training data. Is that a blessing or a curse? Last, the repeating text problem is gone.

Second attempt

Next I tried adding another rule. I added training data to remove please. I created only 5 examples.

{
    "prompt":     "Please click the button.",
    "completion": "Click the button."
},
{
    "prompt":     "To run the job, please select Go.",
    "completion": "To run the job, select Go."
},
{
    "prompt":     "Please enter your username.",
    "completion": "Enter your username."
},
{
    "prompt":     "Please note that the feature is now generally available.",
    "completion": "Note that the feature is now generally available."
},
{
    "prompt":     "For more information, please see the FAQ.",
    "completion": "For more information, see the FAQ."
}

I think there is a way to continue training the first fine-tuned model that I created, but I’ll just create a new one because that seems less prone to pilot error. The new one will have both the it is training data in addition to this new please data.

First I test the please rule by itself:

FIXME: Please configure the app to run on HTTPS only.@STOP!!
Please configure the app to run on HTTPS only.

It didn’t work. 5 examples is probably just not enough. I added 5 more examples, went through the rigmarole of creating a new fine-tuned model, and then tried again:

FIXME: Please configure the app to run on HTTPS only.@STOP!!
Configure the app to run on HTTPS only. Configure the app to run on HTTPS only.
Configure the app to run on HTTPS only. Configure the app to run on HTTPS only.
...

The repeating text problem happened again. But the please was deleted!

Subsequent tests had the same behavior. Another problem that popped up: sometimes the @STOP!! sequence was not removed from the output, like this:

FIXME: Remember to please return to your car.@STOP!! Remember to return to your car.@STOP!

That second @STOP!! at the very end shouldn’t be there.

Finally, I tested the it is and please rules together:

FIXME: Please remember that it is not a toy.@STOP!!
Remember that it's not a toy. STOP: Remember that it's not a toy. STOP:
Remember that it's not a toy. STOP: Remember that it's not a toy. STOP: 
...

Yes, the fine-tuned LLLM “remembered” both rules! OK, I’ve seen enough. Let’s wrap this up.

Cost

I can’t pinpoint the exact cost because I’ve been doing a lot of experimentation of different ideas over the last month but I can tell you that my total cost so far is $1.37. Ballpark estimate, this fine-tuning experimentation cost $1.

Concluding thoughts

I think the “fine-tuning an LLM into a style guide editor” thing can work. I’m going to keep hacking on it. You can follow my progress in this repo: technicalwriting/hank
The fact that I started to see results with only 10 examples for each rule suggests to me that this thing will work robustly if I train it properly on hundreds of examples for each rule, as OpenAI recommends.
When I prepare the training data, I need to make sure that I provide examples on both when to enforce the rule and when not to enforce the rule. For example, the LLM shouldn’t always edit it is down to it's.
This approach seems like it has great potential for collaboration across the technical writing community because the heart of the work is the training data, and that can just be “before” and “after” text in a spreadsheet. The linter approach requires programming and regex. REGEX! The training data is just words. Maybe we will see huge, open source training datasets for style guide rules. You’ll be able to mix and match by just grabbing the training data for only the rules that your style guide follows. Maybe I’ll submit this as a Write The Docs Portland 2023 Writing Day project…
I need to research OpenAI’s models to figure out which one is best suited for this job. I just defaulted to ada because it seemed cheap and well-supported. Also, in the OpenAI Playground I noticed that OpenAI has a beta model that is specifically designed for editing.