ILLATE
Fine-tune rescue

Your OpenAI fine-tune stopped answering? Have it back in 48 hours.

Send the training file you used with OpenAI. Within 48 hours, a small open model trained on it answers through an OpenAI-compatible API, tested on examples it never saw. You pay only once it meets the accuracy we agree on the call.

USD 1,500 fixedPaid on acceptanceYou keep the weights
01 Fit

Who it's for

Good fit

Short, fixed answers

Your fine-tune classified, routed, tagged, scored or moderated, or pulled fixed fields, and you still have the JSONL file you trained it with. These are the tasks a small model matches, and the ones we can prove.

Not a fit

Free-form writing

Summaries, replies and rewrites can't be judged by matching answers, so a 48-hour test can't prove them. We'll say so on the call rather than take your money.

02 Steps

How it works

  1. A 15-minute call

    We look at the task and agree the accuracy bar, usually the level you measured on the retired model. Sign our NDA, or send yours.

    → a one-page order
  2. Send the file

    Share a link to the JSONL you trained with, in OpenAI's chat format or the older prompt and completion format, with poojith@illate.dev only. Not as an attachment.

    → the 48 hours start
  3. The report, within 24 hours

    We train on most of the file and test on the rest, which the model never sees: accuracy with confidence intervals, and results for each category.

    → you see the numbers first
  4. It answers, within 48 hours

    Run it on your own GPU with our vLLM setup, or have us host it behind an OpenAI-compatible endpoint. Your code changes the base URL and the model name.

    → your feature works again
  5. Pay on acceptance

    USD 1,500 once it meets the bar. If it doesn't, you owe nothing.

    → weights are yours
03 Limits

What to know first

Is your fine-tune down?

Book a 15-minute call, or email poojith@illate.dev with "rescue" in the subject. Tell us what the model did and how many examples are in the file.