---
name: pando-map-training
description: Use when the user requests Pando reference-map dataset preparation, diverse macro-map training, checkpoint comparison or ONNX verification.
---

# Pando map training

Use existing Pando command-line tools through the built-in Codex terminal.
This skills-based plugin needs no QGIS, additional desktop app, MCP server,
new virtual environment, background daemon or full application build.

The permanent implementation checkout on this PC is D:/dev/worktrees/ui.
Respect its AGENTS.md and preserve unrelated work. Reuse
build-native/python/Scripts/python.exe and its existing CUDA installation.
Never install a desktop GIS application merely to use this plugin.

## Sources and labels

The canonical owners in tools/reference-analysis/ are acquire-sources.py,
macro.py, corpus.py, train.py, inspect-checkpoint.py and verify-trained-onnx.mjs.
Read their current README and CLI arguments; extend these owners only when a
concrete missing capability requires it. Do not create a parallel training engine.

Resolve external data with tools/report-paths.mjs or tools/report_paths.py and
explicit read/write access. Run code in the app checkout, not the data store.
Keep original maps, annotations, models and reports in external library storage.
Hydrate requested scopes only; reuse source bytes and dependencies.

Prioritize actual world/continent/country maps with different institutions,
production families, scripts, projections, colour/monochrome and scan conditions.
Do not count rescans, translated variants or augmentation as separate originals.
Admit only rights-verified sources and source-bound labels. Unknown dates and
unobserved classes stay unknown. Unlabelled inventory candidates do not train.

Review source/target overlays, legends, text overlaps and disputed marks before
training. Source depiction is not legal attribution. Verify exclusion rectangles
against actual positive/known pixels. For sparse printed strokes, use bounded
known bands where appropriate rather than invented whole-sheet negative truth.
Validate with the current corpus owner before training.

## Train and inspect

Use train.py train --training-only --macro-only --balance-sources with an explicit
admitted corpus, actual prior checkpoint and a new output path. Add
--augment-conditions when scan/photometric diversity is requested; coordinates
and labels stay unchanged. Select epochs, learning rate and encoder freeze for
the actual data. Record attempted batches, successful updates and AMP skips.

Read the initializer reports. Refuse rejected or aborted checkpoints; correct
labels and restart from the last admitted predecessor rather than continuing
weights already trained on rejected targets. Cancel only owned training jobs.

Run focused condition/macro/model checks for actual changes. After real GPU
training, export ONNX and inspect identical unaugmented targets against the actual
baseline with inspect-checkpoint.py. Verify ONNX/PyTorch parity locally using
verify-trained-onnx.mjs in the existing Node WASM runtime. Pass local model and
diagnostic paths; no browser, static server or --base argument is used.

Training-tile fitting is not held-out accuracy. Keep trainingOnly true,
evaluationStatus pending, releaseApproved false and map application not-applied
until independent evaluation and actual device checks justify a separately
authorized release. Keep weights in the shared model scope. Provisional plugin
use requires review and never promotes the model approval fields.

Archive selected external corpus/model/report prefixes using the existing library
storage command. Keep active/best checkpoints and designated evaluation evidence;
clean ended temporary renders and reproducible intermediates. Report blocked
deletion with exact sizes and never bypass execution review. Do not merge main,
publish a model, deploy or rebuild installers without that scope.