LoRA Geometry Diagnostics

LoRA geometry is a learning artifact for the factory. It answers:

Is the skill actually learned by a low-rank effective update?
Which modules move most?
Are DoRA/LoRA updates concentrated enough to try lower rank?
Can we compare failed and successful adapters without rerunning training?

posttrainllm adapters use the TGLA format, so we can inspect effective updates without loading the base model:

python3 scripts/lora_geometry.py runs/<id>/<adapter>.lora --out runs/<id>/lora-geometry.json

The script reports per-entry:

configured rank
realized matrix rank
stable rank
Frobenius norm
spectral norm
whether DoRA magnitude is present

How To Use It

For every meaningful adapter family, compare:

baseline SFT adapter
failed DPO/RL adapter
retry adapter
optionally a rank-truncated or middle-layer-only variant

If a failed adapter has large movement in the wrong modules or a successful adapter has very low stable rank, the next candidate should test lower rank or module targeting before increasing model size.