A100 AYURVEDA TRAINING - 10 CELL SCRIPTS
========================================

For: Gemma-4 12B + 1.69GB clean dataset (971k conversations)
Hardware: A100 (auto-detects 40GB or 20GB MIG)

RUN ORDER:
1. a100_cell_01_install.py   - pip install unsloth, transformers, etc.
2. a100_cell_02_config.py    - auto-detect GPU, set paths, configure hyperparameters
3. a100_cell_03_dataset.py   - load 1.69GB dataset, standardize ShareGPT
4. a100_cell_04_model.py     - load Gemma-4 12B from HuggingFace
5. a100_cell_05_lora.py      - attach LoRA adapter (auto r=64 or 128)
6. a100_cell_06_args.py      - training arguments (cosine LR, adamw_8bit, etc.)
7. a100_cell_07_trainer.py   - initialize SFTTrainer
8. a100_cell_08_train.py     - START TRAINING (~2-4 hours on A100 40GB)
9. a100_cell_09_save.py      - save LoRA adapter (~100-200 MB)
10. a100_cell_10_test.py     - test inference with a sample question

IN JUPYTER:
%run a100_cell_01_install.py
%run a100_cell_02_config.py
...etc...

AUTO-DETECT SETTINGS:
- A100 40GB (>=35GB): batch=2, context=4096, LoRA r=128
- 20GB MIG (>=18GB):  batch=1, context=2048, LoRA r=64
- Low VRAM (<18GB):    batch=1, context=1024, LoRA r=32

DATASET PATH:
Edit DATASET_PATH in cell_02 if your file is not at:
  E:\dataset\gemma4_ayurveda_unsloth_clean.jsonl

OUTPUT:
Saved to OUTPUT_DIR (default: E:\dataset\gemma4_ayurveda_a100_output)
