AssociationAI starter kit: Fine-Tune Your Own Model with Unsloth
===============================================================

WHAT IS IN THIS KIT

qa-template.csv ................ the Q&A spreadsheet template (fictional
                                 teaching rows included; replace with yours)
csv-to-jsonl.html .............. local converter: paste the CSV, get the
                                 JSONL the fine-tune reads, with a PII scan
pii-checklist.txt .............. pre-training data review checklist
unsloth-qlora-starter.ipynb .... the QLoRA starter notebook (run in Colab
                                 or on a local GPU)
eval-questions.txt ............. 10 held-out eval questions + scoring rubric
README.txt ..................... this file

SETUP STEPS (mirrors the tutorial's 8 steps)

1. Decide what may train. Published FAQs, event copy, and de-identified
   member Q&As can train in a cloud notebook. Names, emails, and member
   records train on a local GPU only, or get de-identified first.

2. Write 50 to 200 reviewed Q&A pairs. Start from qa-template.csv and the
   tutorial's training-data studio demo.

3. Convert the spreadsheet. Open csv-to-jsonl.html locally (double-click; no
   network), paste the CSV, read the PII report, and download training.jsonl.

4. Review every pair against pii-checklist.txt. A human reads all of them.
   The scan catches patterns, not names.

5. Hold out 10 pairs for evaluation. Copy them into eval-questions.txt and
   remove them from training.jsonl.

6. Train. Run unsloth-qlora-starter.ipynb: on a free Colab T4 GPU for
   non-sensitive data, or on a local NVIDIA GPU so sensitive data never
   leaves.

7. Score base vs fine-tuned on the 10 held-out questions with the rubric in
   eval-questions.txt. Ship only if the fine-tuned model wins at least 7 of
   10.

8. Pick a deploy route: Colab free T4 (non-sensitive data), local NVIDIA GPU
   (sensitive data never leaves), or dataset now and GPU later.