Hi everyone. I've been using the Mistral platform's API for a while now, mostly for standard chat completions. I'm a cloud admin, so my world is AWS configs, security groups, and cost dashboards 😅.
I want to fine-tune a small model (like Mistral 7B) to understand and generate **specific AWS Terraform patterns**. Think: always outputting modules with security group rules formatted our internal way, with cost-saving tags pre-populated.
Has anyone done something similar? My main questions are:
1. Is the fine-tuning interface on the platform straightforward for a dataset of, say, 100 example Terraform snippets?
2. How do you structure the training data? Is it just a JSONL with `"messages"` containing my examples?
I'm nervous about the cost and the process, but determined to try. Any walkthrough or example of the config would be a huge help.
Here's a sample of the *kind* of snippet I'd train on:
```json
{
"messages": [
{
"role": "user",
"content": "Generate a Terraform AWS security group for a web server on port 80 and 443."
},
{
"role": "assistant",
"content": "resource "aws_security_group" "web_sg" {n name_prefix = "web-sg-"n vpc_id = var.vpc_idnn ingress {n description = "HTTP"n from_port = 80n to_port = 80n protocol = "tcp"n cidr_blocks = ["0.0.0.0/0"]n }nn ingress {n description = "HTTPS"n from_port = 443n to_port = 443n protocol = "tcp"n cidr_blocks = ["0.0.0.0/0"]n }nn tags = {n CostCenter = "platform"n ManagedBy = "terraform"n }n}"
}
]
}
```
Oh, that's a cool use case! I've been thinking about fine-tuning for our internal SQL style guides, so I'm following this closely. Your Terraform idea sounds super practical.
For your data structure question, from what I've read, you're on the right track with the JSONL and "messages" array. But I think you need to include a full conversation history in each record, like starting with a system prompt that says "You are an AWS Terraform expert..." before the user and assistant messages.
How do you plan to handle the system prompt in your training data? I'm still figuring that part out myself.
Good point about the system prompt. Including it as the first message in every "messages" array is indeed the standard approach, and Mistral's platform expects that format.
One caveat for the original poster: for a task like generating formatted Terraform, the model might overfit if your system prompt is too specific about exact formatting rules. It's often better to have a general system prompt about being a helpful AWS expert, and let the bulk of the specific formatting style be taught through the user/assistant example pairs.
For your SQL use case, are you planning to have the model correct queries, or generate new ones from a description? That changes how you'd structure those example conversations.
A hundred examples is nowhere near enough for something as structured as Terraform. You'll burn cash and get a model that regurgitates those exact snippets, failing on any minor variation. The fine-tuning interface is deceptively simple. Upload your JSONL, click a button, watch your credit burn. The real trap is thinking this solves a style guide problem. A linter or a good template engine is cheaper and actually works.
Just saying.
Finally, some sense in this thread. "Deceptively simple" is exactly right - the platform is designed to make you feel like you're doing cutting-edge work with just a few clicks. The real invoice arrives later.
You're spot on about the linter or template engine. It's a classic case of buying a sledgehammer to hang a picture. People get dazzled by the "AI" label and ignore the fact they're spending thousands to automate a process a decent script could handle for the price of a coffee.
But I'll push back slightly on the "nowhere near enough" comment. For certain tasks, a hundred high-quality examples of the *reasoning process* behind a Terraform pattern could work. The failure usually comes from people just feeding raw snippets, expecting the model to magically infer the business logic behind the formatting. That's the real cost burn - paying to train a model on data that lacks the actual intelligence you need it to learn.
Show me the TCO.
Yeah, that's the part I'm worried about. I'm trying to figure out what makes a "high-quality example" vs. just a raw snippet. Is it about adding a user query first, like "How do we format a security group rule for an internal app?" and then the assistant's answer?