ci: fix Gitea runner setup and tighten docs
ci / check (push) Successful in 1m12s

Replace setup-python with setup-uv: the self-hosted Gitea runner lacks the
toolcache and lsb_release dependency, so the job failed before any project
step. uv installs its own Python 3.12; the checks are unchanged.

Align the README and docs with the committed artifacts: quick start serves
from baked artifacts, rebuild steps are separate, and the reference-model
comparison no longer overstates the result. Drop stale external-doc
references and comments that restated the code.
This commit is contained in:
Noel Sabia committed 2026-11-14 09:10:47 +01:00
1 parent 9c7f77f66b
commit 070ecd2a78
13 files changed
+81 -75

No files matched your search

+10 -9
View File
@@ -1,10 +1,10 @@
# Metrics — paper support
All values in this file are generated by `make train` and stored in
`artifacts/metrics.json` (forecast + ablation + per-sensor + small model) and
`artifacts/small_metrics.json` (small-model detail). No number here is typed by
hand: to refresh a table, rerun `make train && make figures` and copy the new
JSON, or quote the JSON directly.
Every value in this file comes from the committed artifacts:
`artifacts/metrics.json` (forecast + ablation + per-sensor + small model),
`artifacts/small_metrics.json` (small-model detail), and
`artifacts/data_audit.json` (raw-input counts). To refresh a table, rerun
`make train && make figures` and quote the new JSON.
## Run provenance
@@ -49,9 +49,10 @@ JSON, or quote the JSON directly.
- WAPE improvement: 17.2 percentage points.
- The mission reference numbers for seasonal naive (5.40 / 44.7% / 0.462) are
reproduced exactly.
- The production row is slightly better than the reference 7-day model
(4.37 / 36.2% / 0.672) because `lag_week` uses the previous *observed* same
weekday/hour across sensor gaps instead of recursion.
- The production row also beats the mission reference 7-day no-weather model
(4.37 / 36.2% / 0.672): weather supplies most of the gain, and `lag_week`
uses the previous *observed* same weekday/hour across sensor gaps instead of
recursion.
## Ablation
@@ -174,7 +175,7 @@ Per-sensor small-model holdout:
- The small classifier as the production forecast. It is a demo artifact with
six features and a per-window p75 target.
- Tuned hyperparameters. The model configuration is fixed; the 28-day holdout
is the honest signal, not a search result.
is the held-out result, not a search result.
## Reproduce
+3 -5
View File
@@ -1,10 +1,8 @@
# Small model — busy-hour classifier
This is the small, fast, explainable model that powers the "train new data"
demo and the "I coded an AI" artifact. It is deliberately separate from the
production 7-day forecaster: six cheap features, a linear model, a time-based
split, and under a second of fitting. It is not the model behind the dashboard
forecast.
A logistic regression that labels an hour `busy` when footfall exceeds that
sensor's p75. It powers the `model=small` retraining demo, not the dashboard
forecast: six cheap features, a time-based split, and under a second to fit.
Every number below comes from the baked artifact
`artifacts/small_metrics.json` (version `logit-20261114-052723`, trained
+3 -3
View File
@@ -60,9 +60,9 @@ Input: 36,846 density rows, 8 sensors, 2026-09-08T23:53Z → 2026-11-07T23:51Z.
This is a different window from the baked full-window model in
`docs/small-model.md`, so the numbers are not directly comparable: the p75
target and the 80/20 split are recomputed inside the 60-day upload. The
low recall illustrates the honest limit noted in the small-model doc — a short,
recent window with a different threshold is a different classification problem.
target and the 80/20 split are recomputed inside the 60-day upload. The low
recall illustrates the limit noted in the small-model doc: a short, recent
window with a different threshold is a different classification problem.
Post-conditions verified: