24679-hand-middle-finger-autogluon-regressor

Middle-finger length regressor (AutoGluon) — 24-679 Fall 2026

Purpose

Predicts a person's right-hand middle-finger length in centimetres from the lengths of the other four fingers and a recorded Female / Male category. Built for the 24-679 AutoML homework as an exercise in tabular AutoML, provenance-aware splitting, and model documentation. It is a classroom demonstration on a 30-person convenience sample and is not a tool for anthropometric, medical, forensic, or demographic use.

Data origin and splits

• Source dataset: cmuchancel/hw1-tabular-hand-data (commit 511326a09c89d23a3698deccb1119f3ae3a4e066), prepared by a classmate for Homework 1. 30 CMU students each measured their own right-hand finger lengths (fingertip to the web between fingers) with the same ruler, in centimetres.
• Published splits (kept fixed): train = 21 original hands + 504 synthetic rows (525 rows), validation = 4 original hands, test = 5 original hands. The dataset's own split seeds were 24679 (train/holdout) and 24680 (validation/test).
• Synthetic rows (training only) come from the dataset's documented augmentation: additive Gaussian jitter, multiplicative scaling (×0.98–1.02) of the four predictor lengths, and regression Mixup between two original training hands (weight 0.60–0.90). Every synthetic row descends only from original training hands, which the notebook asserts.
• Sex category counts:

Features and target

Excluded from the feature matrix: source_id, parent_id, second_parent_id, augmentation, is_augmented, mix_weight.

Preprocessing

No manual scaling or encoding. Numeric columns are cast to float64 and checked to be finite and positive; Female / Male is passed as a pandas category. AutoGluon's AutoMLPipelineFeatureGenerator handles type inference and per-model encoding (e.g. the two-level category becomes a boolean). Inputs to predict must use the exact column names above.

Training setup and search space

• Framework: AutoGluon TabularPredictor 1.6.1, problem_type="regression", eval_metric="root_mean_squared_error", presets="medium_quality", tuning_data = the published validation hands, num_bag_folds=0, num_stack_levels=0 (no internal re-splitting of related synthetic rows).
• Stage 1 — two training variants, 300 s each: originals_only (21 rows) and originals_plus_synthetic (525 rows), default medium_quality model zoo (LightGBM ×3, CatBoost, XGBoost, RandomForest, ExtraTrees, KNN ×2, NeuralNetTorch, NeuralNetFastAI, weighted ensemble) plus AutoGluon's linear model LR. Backends available in this run: {'lightgbm': '4.6.0', 'xgboost': '3.3.0', 'catboost': '1.2.10', 'torch': '2.11.0+cpu', 'fastai': '2.8.7'}.
• Stage 2 — random hyperparameter search on originals_plus_synthetic (the better stage-1 variant): 20 trials per family, {'num_trials': 20, 'scheduler': 'local', 'searcher': 'random'}, 600 s budget, over
• GBM: num_leaves ∈ integers 4–64, learning_rate ∈ 0.01–0.3 (log scale), feature_fraction ∈ 0.5–1.0, min_data_in_leaf ∈ integers 2–30, lambda_l2 ∈ 0.001–10.0 (log scale)
• XGB: max_depth ∈ integers 2–8, learning_rate ∈ 0.01–0.3 (log scale), subsample ∈ 0.5–1.0, colsample_bytree ∈ 0.5–1.0, min_child_weight ∈ integers 1–10
• CAT: depth ∈ integers 3–8, learning_rate ∈ 0.01–0.3 (log scale), l2_leaf_reg ∈ 1.0–10.0 (log scale)
• LR: 5 fixed configurations: {'penalty': 'L2', 'C': 0.01}; {'penalty': 'L2', 'C': 0.1}; {'penalty': 'L2', 'C': 1.0}; {'penalty': 'L2', 'C': 10.0}; {'penalty': 'L2', 'C': 100.0}
• RF: 3 fixed configurations: {'n_estimators': 300, 'max_features': 0.5, 'min_samples_leaf': 1}; {'n_estimators': 300, 'max_features': 0.5, 'min_samples_leaf': 3}; {'n_estimators': 300, 'max_features': 1.0, 'min_samples_leaf': 5}
• Selection rule: lowest validation RMSE across all candidates of all runs. Test hands were scored once, afterwards.
• Seeds: Python / NumPy / PyTorch seed 24679; AutoGluon model seed 0 (default); seed-sensitivity refits use seeds [0, 1, 2, 3, 4].
• Fit times (s): {'originals_only': 14.9, 'originals_plus_synthetic': 30.6, 'hpo_originals_plus_synthetic': 76.3}

Best model and hyperparameters

• Run: hpo_originals_plus_synthetic — model: WeightedEnsemble_L2 — trained on 525 rows.
• Validation RMSE: 0.235 cm.

{
  "model": "WeightedEnsemble_L2",
  "model_type": "WeightedEnsembleModel",
  "val RMSE (cm)": 0.23503504300766215,
  "hyperparameters": {
    "ag_args_ensemble": {
      "save_bag_folds": true
    }
  },
  "ensemble_weights": {
    "LightGBM/T4": 0.26666666666666666,
    "LinearModelC100.0": 0.7333333333333333
  },
  "base_models": [
    {
      "model": "LightGBM/T4",
      "model_type": "LGBModel",
      "val RMSE (cm)": 0.4106743419165846,
      "hyperparameters": {
        "num_leaves": 50,
        "learning_rate": 0.1477317633429356,
        "feature_fraction": 0.9818313802505146,
        "min_data_in_leaf": 25,
        "lambda_l2": 0.03417952912061012,
        "seed": 0
      }
    },
    {
      "model": "LinearModelC100.0",
      "model_type": "LinearModel",
      "val RMSE (cm)": 0.265527465630884,
      "hyperparameters": {
        "penalty": "L2",
        "C": 100.0
      }
    }
  ]
}

Metrics

Validation and test scores (RMSE and MAE in cm)

• Test RMSE bootstrap 95% CI: 0.112–0.266 cm; test MAE 95% CI: 0.069–0.258 cm (2000 resamples of the 5 test hands).
• Seed sensitivity (best single model LinearModelC100.0, seeds [0, 1, 2, 3, 4]): test RMSE 0.278 ± 0.000 cm (mean ± sd).
• Scale reference: the standard deviation of the target among training originals is 0.496 cm, and the mean-prediction baseline's test RMSE is 0.339 cm.

Per-hand test predictions

Permutation feature importance (training originals, 20 shuffles; RMSE units)

The full validation-vs-test leaderboard for every candidate is in leaderboard_all_candidates.csv; the machine-readable version of everything above is in metrics.json.

Limitations and ethical notes

• Tiny sample. 30 hands in total, of which 4 select the model and 5 test it: one test hand shifts RMSE substantially, and selection on 4 validation hands is noisy (see the val-vs-test table). Synthetic rows add no new people.
• Not representative. A convenience sample of CMU students, heavily male (see counts above); measurement error from self-measurement with a ruler is likely of the same order as the model's errors.
• Sex category. Female / Male is a recorded label used only as a predictor. Nothing here should be read as a claim about biological differences, and the model must not be used to infer sex or identity from hand measurements.
• Scope. Right-hand, adult-student measurements in cm; predictions outside the observed ranges are extrapolations.
• No personal identifiers are stored; source_ids are provenance keys only.

License

The underlying dataset assigns no license and is shared for 24-679 coursework; this model is released under the same classroom-use-only terms (license: other). Check course permissions before any reuse or redistribution.

Hardware and compute budget

Linux-6.6.122+-x86_64-with-glibc2.39, Python 3.13.15, 2 CPU cores, 13.6 GB RAM, GPU: none (CPU runtime). Budgets: 300 s per stage-1 variant, 600 s for stage 2, 120 s per seed-check refit; actual fit times are listed above. Run started 2026-09-21T00:50:36+00:00. Library versions: AutoGluon 1.6.1, scikit-learn 1.7.2, pandas 2.2.3, NumPy 2.1.3, PyTorch 2.11.0+cpu.

How to use

import zipfile, huggingface_hub, pandas
from autogluon.tabular import TabularPredictor

path = huggingface_hub.hf_hub_download("shanexf/24679-hand-middle-finger-autogluon-regressor", "autogluon_predictor_dir.zip")
zipfile.ZipFile(path).extractall("predictor_dir")
predictor = TabularPredictor.load("predictor_dir")
hand = pandas.DataFrame([{"Thumb Size": 6.2, "Index Finger Size": 7.6, "Ring Finger Size": 7.5,
                          "Pinky Size": 6.0, "Female / Male": "Male"}])
print(predictor.predict(hand))

autogluon_predictor.pkl is a cloudpickle of the same predictor and depends on the files in the ZIP; use the same AutoGluon version listed above.

AI usage disclosure

Claude (Anthropic) was used to adapt the 24-679 lecture AutoML notebook to this dataset and to draft the model-card template. The author ran the notebook end to end, checked the outputs, and edited the text. The hand measurements were collected by a classmate from CMU students and were not generated by AI.

Author

shanexf — 24-679, Carnegie Mellon University, Fall 2026.