Function train_and_save_model

uncommon_route/router/classifier.py:168–207

Train model from JSONL data — keyword-free.

(data_path: str, out_path: str | None = None)

def train_and_save_model(data_path: str, out_path: str | None = None) -> None:
    """Train model from JSONL data — keyword-free."""
    import json

    cases = []
    with open(data_path, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if line:
                cases.append(json.loads(line))

    feature_sets: list[tuple[dict[str, float], str]] = []
    model = ScriptAgnosticClassifier(use_ngrams=True)

    for case in cases:
        prompt = case["prompt"]
        feature_text = _merge_feature_text(prompt, case.get("system_prompt"))
        struct_dims = extract_structural_features(feature_text)
        structural_scores = {d.name: d.score for d in struct_dims}
        unicode_blocks = extract_unicode_block_features(feature_text)

        features = model._build_features(
            structural_scores,
            unicode_blocks,
            keyword_scores=None,
            prompt=feature_text,
        )
        normalized_tier = model._normalize_tier_label(case["expected_tier"])
        if normalized_tier is not None:
            feature_sets.append((features, normalized_tier))

    model.train(feature_sets, epochs=20)

    save_to = Path(out_path) if out_path else Path(__file__).parent / "model.json"
    model.save(save_to)
    print(f"Trained on {len(cases)} cases, saved to {save_to}")

    correct = sum(1 for feats, tier in feature_sets if model.predict(feats)[0] == tier)
    if feature_sets:
        print(f"Training accuracy: {correct}/{len(feature_sets)} ({correct / len(feature_sets) * 100:.1f}%)")


# ─── Trivial Detection (structural only, no keyword lists) ───