Exam AI: LoRA fine-tuning on Kaggle's free GPUs

One permissively licensed small base model, one QLoRA adapter per exam family, trained only on material this project has the right to use. This is phase 2 of ops/plan/exam-llm.md (task T-123); phase 3 (T-124) publishes the winning adapter by hand.

File	What it is
exam-ai-lora.ipynb	The Kaggle notebook: license check, data build, training, evaluation, release gate.
licenses.json	Every base model and dataset with its license, the URL verified against, and the date.

Run it (free)

1. Kaggle -> New Notebook -> File -> Import Notebook -> upload exam-ai-lora.ipynb.
2. Accelerator: GPU T4 x2 (or P100). Internet: on. ~45–90 minutes of the ~30 free GPU hours a week.
3. Attach this repository as a Kaggle dataset (upload the folder whole) or let the notebook clone the public GitHub repository. No API keys, no paid calls.

The release rule (enforced, not promised)

• The *-eval-* files (questions/ee-eval-*.json, 50 questions, T-122) are held out: the notebook scores the base model and each adapter on them and never trains on them (scripts/lora_data.py:split_train_eval).
• An adapter is copied to /kaggle/working/release/<family>/ only if it beats the base model on those questions; otherwise the notebook raises after writing results.json and releases nothing.

• scripts/lora_release_gate.py recomputes every recorded score from the predictions and the answer key, so a hand-edited results file cannot claim a win. Run it locally:

  python3 scripts/lora_release_gate.py --registry notebooks/lora/licenses.json --results results.json

• A family without a T-122 evaluation slice gets no release (its adapter stays a candidate). Exact dates, cutoffs and vacancy counts are never learned as facts: the format taught is "Not sure - check the official notification" with the official link.

Licenses (verified 2026-09-28 against the pages linked in licenses.json)

Used	Model / dataset	License
base	Qwen/Qwen2.5-1.5B-Instruct	Apache-2.0
alternative	HuggingFaceTB/SmolLM2-1.7B-Instruct	Apache-2.0
rejected	PhysicsWallahAI/Aryabhata-1.0	CC BY-NC 4.0 (blocks commercial use)
data	this list's pages, modules, formula sheets	CC BY-SA 4.0
data	this list's original questions	CC BY-SA 4.0
replay	HuggingFaceH4/ultrachat_200k (capped sample)	MIT

The adapters are released under CC BY-SA 4.0, matching the content they learn from; attribution "Awesome Indian Exams by Rajon Das and contributors".

Outputs

/kaggle/working/results.json (scores, per-question predictions), /kaggle/working/adapters/<family>/ (candidates), /kaggle/working/release/<family>/ (adapter + MODEL_CARD.md, only when the gate passes).