SOURCE-DOMAIN STUDY · v0.1
Wine Domain Transfer
Compare two small source-domain regressors on fixed red and white wine holdouts.
Each training pool has 923 rows and uses the same update budget. Every source-domain model selects its checkpoint on its own validation set. Input duplicates stay within one split across both colors.
Known public data, not a blind benchmark. Sensory-score points are not probabilities, wine certification or a health recommendation.
What changes across domains?
On narrow screens, scroll the tables to view all columns.
| Test domain | Red-trained MAE | White-trained MAE | Red fit-mean MAE | White fit-mean MAE | Full test count per seed |
|---|
Loading fixed study…
Equal row/update budgets do not isolate a causal shift: quality distributions, normalizers, feature support, duplicate weighting and own-domain validation differ. Exact input groups are not producer IDs or proof of independence.
Compare one input
Use source numeric scales. The bundled files do not specify individual chemical units; none are assumed here. Accepted bounds are source-global marginal ranges, which do not establish joint support or physical validity.
The fixed default red:0002 is the lowest red-test ordinal, chosen before model test predictions. Other presets use the lowest source ordinal for each role. No error-based example selection.
Exact-schema numeric JSON input
All 11 original feature names are required, with numeric values only and no extra keys. A JSON edit selects JSON mode. Changing the seed rechecks that mode; invalid JSON cannot revive earlier field values.
Checking model and catalog hashes…
| Training source | Seed | Selected epoch | Prediction (score points) | Exact group seen in fit | Outside fit marginals |
|---|
Outputs are not rounded to classes or clipped to 0–10. Model confidence, tasting quality and future-domain performance are not established. A numeric match to a source vector does not authenticate a measurement.
Complete current report JSON
Full experiment evidence
Six fresh 11→24→1 neural fits, 313 parameters each; 600 total epochs, 9,000 updates. Three initial sets are shared within paired seeds. Each model uses 923 fit rows and 335 own-source validation rows. Two fit-mean constants are analytic baselines.


Complete result summary · All 10,584 method predictions · All paired comparisons · Every in-domain failure · All 6,497 source records and roles
Source and limits
UCI Wine Quality, Cortez, Cerdeira, Almeida, Matos and Reis (2009), DOI 10.24432/C56S3T, CC BY 4.0. Red 1,599 + white 4,898 original records, 5,318 exact input groups, including two cross-color groups. All source labels/distributions were audited before protocol; this is a known-data experiment.
The target is a sensory score on a stated 0–10 scale; original names describe the median of at least three expert evaluations. The original paper is cited via source metadata, not reproduced. Original code: Renchao Wu, Apache-2.0. Source data, derived weights, results and figures carry CC BY 4.0 attribution.
This page computes locally with native JavaScript. No external model API, analytics or browser storage. Edits are lost on refresh. Deliberate downloads include all input values, all model outputs and provenance; hashes establish byte consistency, not authenticity.
Reproduction and limitations · Fixed protocol · Research sources · License scope