Add TF-IDF + LinearSVC TweetEval sentiment example tuned with Optuna (with smoke test & README)

examples/sklearn/svm_tfidf_tweeteval_sentiment.md

@@ -0,0 +1,9 @@
+# SVM TF-IDF Sentiment (TweetEval) tuned with Optuna
+
+This example tunes a TF-IDF + LinearSVC pipeline on the TweetEval **sentiment** dataset (3 classes: negative, neutral, positive) using Optuna.
+
+**Run:**
+```bash
+python examples/sklearn/svm_tfidf_tweeteval_sentiment.py --n-trials 20 --max-train 20000
+
+

examples/sklearn/svm_tfidf_tweeteval_sentiment.py

@@ -0,0 +1,75 @@
+from __future__ import annotations
+import argparse, random
+import numpy as np
+import optuna
+from datasets import load_dataset
+from sklearn.metrics import classification_report, f1_score
+from sklearn.pipeline import make_pipeline
+from sklearn.svm import LinearSVC
+from sklearn.feature_extraction.text import TfidfVectorizer
+
+def set_seed(seed: int) -> None:
+    random.seed(seed); np.random.seed(seed)
+
+def load_tweeteval_sentiment(max_train: int, seed: int):
+    ds = load_dataset("tweet_eval", "sentiment")
+    def to_xy(split):
+        X = [r["text"] for r in ds[split]]
+        y = [int(r["label"]) for r in ds[split]]
+        return X, y
+    X_train_all, y_train_all = to_xy("train")
+    X_valid, y_valid = to_xy("validation")
+    X_test, y_test = to_xy("test")
+    if max_train and 0 < max_train < len(X_train_all):
+        rng = np.random.default_rng(seed)
+        idx = rng.choice(len(X_train_all), size=max_train, replace=False)
+        X_train = [X_train_all[i] for i in idx]; y_train = [y_train_all[i] for i in idx]
+    else:
+        X_train, y_train = X_train_all, y_train_all
+    return (X_train, y_train), (X_valid, y_valid), (X_test, y_test)
+
+def build_pipeline(trial: optuna.Trial):
+    max_features = trial.suggest_int("tfidf_max_features", 10000, 60000, step=5000)
+    min_df = trial.suggest_float("tfidf_min_df", 1e-5, 5e-3, log=True)
+    ngram_high = trial.suggest_int("tfidf_ngram_high", 1, 2)
+    sublinear_tf = trial.suggest_categorical("tfidf_sublinear_tf", [True, False])
+    C = trial.suggest_float("svm_C", 1e-2, 10.0, log=True)
+    loss = trial.suggest_categorical("svm_loss", ["hinge", "squared_hinge"])
+    class_weight = trial.suggest_categorical("svm_class_weight", [None, "balanced"])
+    vec = TfidfVectorizer(max_features=max_features, min_df=min_df,
+                          ngram_range=(1, ngram_high), sublinear_tf=sublinear_tf, norm="l2")
+    svm = LinearSVC(C=C, loss=loss, class_weight=class_weight, random_state=0)
+    return make_pipeline(vec, svm)
+
+def objective(trial: optuna.Trial, X_train, y_train, X_valid, y_valid):
+    pipe = build_pipeline(trial)
+    pipe.fit(X_train, y_train)
+    y_pred = pipe.predict(X_valid)
+    return 1.0 - f1_score(y_valid, y_pred, average="macro")
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--n-trials", type=int, default=10)
+    ap.add_argument("--max-train", type=int, default=15000)
+    ap.add_argument("--seed", type=int, default=42)
+    args = ap.parse_args()
+    set_seed(args.seed)
+    (Xtr, ytr), (Xva, yva), (Xte, yte) = load_tweeteval_sentiment(args.max_train, args.seed)
+    study = optuna.create_study(direction="minimize",
+                                sampler=optuna.samplers.TPESampler(seed=args.seed),
+                                pruner=optuna.pruners.MedianPruner(n_startup_trials=5))
+    study.optimize(lambda t: objective(t, Xtr, ytr, Xva, yva),
+                   n_trials=args.n_trials, show_progress_bar=True)
+    best = study.best_trial
+    print("\nBest validation (1 - macroF1):", best.value)
+    print("Best params:", best.params)
+    # retrain on train+valid, evaluate on test
+    pipe = build_pipeline(best)
+    pipe.fit(Xtr + Xva, ytr + yva)
+    y_pred = pipe.predict(Xte)
+    print("\n=== Test set ===")
+    print("macro-F1:", f1_score(yte, y_pred, average="macro"))
+    print(classification_report(yte, y_pred, digits=4))
+
+if __name__ == "__main__":
+    main()

tests/test_svm_tfidf_tweeteval.py

@@ -0,0 +1,11 @@
+import os, sys, subprocess
+
+def test_cli_help_prints_without_running():
+    script = os.path.join("examples","sklearn","svm_tfidf_tweeteval_sentiment.py")
+    proc = subprocess.run([sys.executable, script, "--help"],
+                          capture_output=True, text=True)
+    assert proc.returncode == 0, proc.stderr
+    out = proc.stdout
+    # Should mention our CLI arguments
+    assert "--n-trials" in out
+    assert "--max-train" in out