Just wrapped up my first fine-tuning project on OpenPipe using the Llama 3.1 8B Instruct model and wanted to share a real-world cost breakdown. I was migrating some old Salesforce CPQ logic into a custom assistant, and needed it to understand our specific product bundling terminology.
My training dataset was 1,247 examples in OpenAI's `jsonl` chat format. Each example was a conversation simulating our internal Q&A style. Here's a snippet of the config I used for the OpenPipe CLI:
```json
{
"model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
"training_data": "cpq_qa.jsonl",
"epochs": 3,
"learningRate": 2e-5
}
```
The total cost for the training run came to **$4.67**. This covered the entire process, from uploading the data to receiving the fine-tuned model ID. I was pleasantly surprised by the clarity—no separate charges for compute hours or storage during training.
A few practical observations:
* The cost scales linearly with dataset size and epochs. A teammate ran a similar job with ~500 examples for 2 epochs and paid just under $2.
* No hidden inference costs for the training itself. You only pay for inference when you start calling your new model.
* The web UI gives you a cost estimate before you commit, which is great for budgeting.
Has anyone else fine-tuned the 70B or 405B versions? I'm curious how the cost scales for the larger models, and if the performance jump is worth it for complex data mapping tasks. I'm planning a Zapier integration assistant next.
hth