| --- |
| license: mit |
| library_name: scikit-learn |
| pipeline_tag: tabular-classification |
| tags: [tabular, cpu, tiny, cog, app-nz] |
| --- |
| |
| # mojo-tabular |
|
|
| Tiny CPU tabular classifiers (one per dataset), total 12 MB, no GPU. Trained with scikit-learn; |
| `mojo-` is the planned optimization path, not a Mojo implementation yet. |
|
|
| ## Results (ROC-AUC, seed 42, 60/20/20 stratified split, selection on validation only) |
|
|
| | dataset (OpenML) | val | test | untuned best baseline test | MB | |
| |---|---|---|---|---| |
| | adult | 0.929 | 0.9282 | 0.9272 | 2.64 | |
| | credit-g | 0.8294 | 0.7986 | 0.7958 | 0.00 | |
| | bank-marketing | 0.9354 | 0.9384 | 0.9375 | 2.67 | |
| | phoneme | 0.9523 | 0.9546 | 0.9554 | 5.50 | |
| | diabetes | 0.8528 | 0.8001 | 0.8031 | 1.09 | |
|
|
| Search: 16 random HistGradientBoosting configs + logistic regression + extra trees; top-3 averaged |
| when validation improves. Single split, single seed: small sets (credit-g n=1000, diabetes n=768) have |
| large validation/test gaps and are noisy. No comparison against XGBoost/CatBoost/TabPFN has been run, |
| so **no frontier claim is made**. Adding tuned XGBoost 3.4.1 and LightGBM 4.7.0 (6 configs each) to the pool changed test AUC by at most +0.001 (adult 0.929, bank-marketing 0.9386), i.e. these results sit at the plateau of strong boosted trees; the shipped models use sklearn only to keep the image small. CatBoost and TabPFN were not compared (CatBoost hung under the 4 GiB address-space limit). Data: public OpenML sets (license listed as Public). |
| Reproduce: `train/fetch_tabular.py`, `train/tabular_tune.py` (`--gbdt` needs xgboost/lightgbm), artifact hashes in `results/tuned.json`. |
|
|
| ## Use |
|
|
| ```python |
| import joblib, pandas as pd |
| b = joblib.load("models/adult.joblib") # needs scikit-learn==1.8.0 |
| p = sum(m.predict_proba(df[b["columns"]])[:, 1] for m in b["members"]) / len(b["members"]) |
| ``` |
|
|
| ## Deploy on app.nz (Cog, CPU) |
|
|
| ```bash |
| cd cog && python3 -m unittest discover -s tests # contract test, no network |
| docker build -t ghcr.io/lee101/mojo-tabular:latest . # 381 MB (python:3.12-slim), weights baked |
| docker run --rm -p 5000:5000 --memory 512m ghcr.io/lee101/mojo-tabular:latest # ~80 MB RSS, ready in <10 s |
| ``` |
|
|
| [](https://app.nz/deploy?image=ghcr.io/lee101/mojo-tabular:latest&name=mojo-tabular&hardware=cpu&idleSeconds=60) |
|
|
| `POST /predictions {"input": {"dataset": "adult", "rows": "[{...}]"}}` returns a JSON list of probabilities |
| (max 1000 rows). Do not use the default `cog build` image: its apt layer alone is 1.78 GB (2.33 GB total). |
| Links: [GitHub](https://github.com/lee101/mojo-tabular), [Hugging Face](https://huggingface.co/lee101/mojo-tabular). |
| Status: no registry image is published yet. The GHCR push needs a token with `write:packages`, and app.nz's remote builder currently runs in simulated mode (`APPNZ_BUILDER` unset), so the one-click badge will not resolve until an image is pushed. Build locally with the command above. |
|
|
| ## Limitations |
|
|
| Trained on public benchmark data only; not validated for credit, hiring or medical decisions. |
|
|
