Building a fine-tuned code review bot for an open-source Golang project

AI Projects
, ,
1

Hello everyone.

So I’ve been learning and reading everything about LLMs and trying to wrap my head around it all.

I’m fairly confident in knowing how to train and fine-tune once I have a dataset. What I still don’t know and I feel like the coursera course also glanced over this by using their own dataset, is dataset generation.

I have an idea for a project of fine-tuning bigcode/starcoder · Hugging Face for a more specific code suggestion on a relatively large golang codebase.

Then, would write a bot to use it for codde reviews or a coding buddy basically. I know huggingface had a post like, create your own copilot, but that used a dataset generated from Jupyter notebooks.

So my question is… how do I generate a dataset using Golang code, without handcrafting prompts? What’s the format that the dataset needs to be in? Can I write a program that generates it for me? I looked at some similar datasets on huggingface, but they all look like they have been handcrafted which is insane. :smiley: But if that’s what it takes I’ll do it, but I’m just not sure if there is anything better out there?

thanks!

2

Would this be a valid dataset format for fine-tuning?

[
    {
        "text": "func init() {\n\tutilruntime.Must(clientgoscheme.AddToScheme(scheme))\n\tutilruntime.Must(corev1.AddToScheme(scheme))\n\tutilruntime.Must(deliveryv1alpha1.AddToScheme(scheme))\n\tutilruntime.Must(v1.AddToScheme(scheme))\n\t//+kubebuilder:scaffold:scheme\n}",
        "metadata": {
            "file_name": "main.go",
            "file_path": "./cmd/main.go",
            "line_number": 50,
            "func_name": "init"
        }
    },
    {

Basically the text would be the function and then some metadata.