Datadog/Toto-2.0-22m is the second-smallest variant in the Toto 2.0
family. Tighter confidence bands and lower MAE for ~5Γ the parameter
count, still fits comfortably in CPU memory and runs sub-second per
metric. Patch size is read from model.config at runtime so the
context-trim / pad logic adapts automatically.
Updates the subtitle on the page and the inference doc to reflect
the variant change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
docs: document the hybrid cadence (5-min display, hourly inference)β
Update the Toto inference spec to reflect the current setup: the chart
shows 5-min Ecowitt actuals from the local SQLite archive, but the
model itself sees the same series resampled to hourly. Drop the stale
dropdown references β the cadence and horizon knobs were removed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Zoom view uses 5-min cadence (Ecowitt's actual storage granularity)β
Probe confirms /device/history at cycle_type=5min returns 300-second
gaps on this station; the GW3000B is uploading at ~1-min cadence and
Ecowitt buckets to 5 min for the 90-day tier. Switch the zoom view from
30-min to 5-min for both history and forecast:
- 36 h history at 5-min spacing = 432 points β trimmed to 416 (13
Toto patches of 32).
- 12 h horizon at 5-min spacing = 144 forecast steps per metric.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Document scoreboard methodology in-app and in the specβ
- docs/toto-inference.md: expand the scoreboard section β input tables,
the 'latest pre-target forecast' rule that prevents either model from
peeking, the MAE formula on p50 vs actual, and explicit caveats
(point-only score, all horizons folded together).
- app.py: new 'How the scoreboard is calculated' accordion above the
existing 'How the forecast is made' one, summarising the same rules.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Document Toto inference: full spec in docs/ + in-app footnoteβ
- docs/toto-inference.md: precise spec of how the demo makes forecasts β
model id and params, input shape and resampling, context length and
padding rule, prediction horizon math, distribution-to-quantile
semantics (analytical, not Monte-Carlo), inference cadence, and the
scoreboard / persistence layer.
- app.py: collapsible 'How the forecast is made' Accordion at the bottom
of the Space with the brief version, linking to the full doc on the
Space's repo browser.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add Ecowitt API client and incremental SQLite archiveβ
- src/ecowitt.py: Ecowitt Cloud API v3 client (info, real_time, history)
with typed errors and a history β DataFrame helper.
- src/storage.py: SQLite schema and idempotent upsert keyed on
(cycle_type, channel, metric, ts_unix); handles deeply nested payloads.
- src/sync.py: incremental sync that resumes from MAX(ts_unix) per
cycle_type, chunks long ranges, and backs off on rate-limit.
- docs/plan.md: high-level project plan.
- docs/ecowitt-notes.md: API findings (valid cycle_types, call_back
rejecting 'all' for history, rate-limit code -1) and run-it-yourself docs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>