Kaggle is not "free magic." It is a quota game with real failure modes: weekly GPU hours, TPU ceilings, HF rate limits, and preempted notebooks.
How we use it
- Private corpora stay on Kaggle-private paths when needed; audited-clean lanes can use other free GPUs
- LoRA lanes stay separated (personality vs music vs historian vs NSFW quarantine)—never melt them into one poisoned adapter
- Artifacts come home — adapters that never leave the cloud are cosplay
Lessons from the frontier week
- Transformers major versions break trainers; pin and probe
- TPU NaNs are often LR + precision, not "bad data"
- Colab preemption requires checkpoint discipline
FAQ
Are Kaggle medals the goal?
No. Adapters that serve are the goal.
Why publish this?
Because "we train models" without process is influencer fog.
What's the privacy rule?
Private directives and family data never ride random third-party GPUs.
Where do models run after training?
Local serving paths (lane servers) with proof before anything is marked live.
Sovereign stack field notes
Local AI, multi-desk news, and infrastructure you actually own.