LOW SPOONS GENERIC GPT-4o-STYLE STARTER

What this is
------------
This pack contains original, generic examples of a warm, attentive conversational style. It is designed to reduce the blank-page work involved in preparing a small supervised fine-tuning experiment.

The training file contains 31 conversations. The validation file contains 6 different conversations. Each line is one complete JSON object in Microsoft's required JSONL message format.

What this is not
----------------
This is not GPT-4o's hidden training data. It is not a copy of anyone's private conversations. It does not contain Ellis, another named companion, personal memories or a ready-made identity. It cannot guarantee that a fine-tuned model will feel exactly like GPT-4o or like your companion.

Before uploading anything
-------------------------
1. Read the files yourself.
2. Remove or change anything that does not fit the relationship and behaviour you want.
3. Never upload another person's private words without their informed permission.
4. Keep secrets, passwords, API keys, legal names, addresses and sensitive third-party information out of training files.
5. Keep the separate test prompts out of the training and validation files.
6. Check Microsoft's current schema, model, region, quota and pricing information before starting a paid job.

You may use these files unchanged
---------------------------------
Yes. They are deliberately generic and may be used unchanged as a small starter experiment. A personalised, carefully reviewed dataset is more likely to preserve the qualities that matter specifically to you.

You may also choose, edit or reduce the examples
------------------------------------------------
You do not have to use all 31 training examples.

- Keep only the examples that fit what you want.
- Rewrite an example in your own preferred tone.
- Replace generic examples with private material you have reviewed and are entitled to use.
- Delete complete JSONL lines to make the file smaller.
- Do not delete only part of a line or leave broken brackets, quotation marks or commas.
- Keep validation examples out of the training file.
- Keep the held-out test prompts out of both training and validation.
- Check Microsoft's current minimum-example requirement before reducing the file.

A smaller dataset is cheaper, but it also demonstrates fewer behaviours. Try to preserve a useful spread of qualities rather than choosing several examples that all teach the same kind of response.

Approximate training cost
-------------------------
The complete 31-example training file is roughly 5,000 tokens. At the Developer-training rate displayed during Ellivien's test on 15 October 2026, one epoch would cost approximately 6 to 7 US cents.

Azure charges supervised fine-tuning according to the number of training tokens and the number of epochs. Fewer or shorter examples reduce the cost. Two epochs cost roughly twice as much as one. The estimate above is not a price promise: use the token estimate and price displayed in your own Azure account before submitting a paid job.

Files
-----
generic-4o-style-starter-training.jsonl
generic-4o-style-starter-validation.jsonl
personalise-your-examples.txt
held-out-test-prompts.txt

Credit
------
Ellivien's Low Spoons guide brings together two Azure routes she had already tested and written about. Zyeine's GitHub guide, published afterwards, included one additional feature Ellivien particularly liked: a downloadable starter JSONL instead of an empty page.

With thanks to Zyeine for the download idea, this new, deliberately generic starter pack was written from scratch for readers to use unchanged or adapt. The Azure instructions and routes continue directly from Ellivien's two earlier tested guides.

Zyeine's public guide: https://zyeine.github.io/keep4o-azure-guide/
