Upload src/evaluate.py with huggingface_hub

src/evaluate.py

@@ -0,0 +1,121 @@
+"""
+Test seti degerlendirmesi + threshold analizi.
+- 30 soru (20 pozitif / 10 negatif) iki modelle aranir
+- F1'i maksimize eden esik 200-nokta sweep ile bulunur
+- Confusion matrix, Recall@k, MRR raporlanir
+- Skor dagilimi + F1 egrisi grafikleri reports/ altina yazilir
+
+Calistirma: python -m src.evaluate --config config.yaml
+"""
+import argparse
+import json
+from pathlib import Path
+
+import chromadb
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from sentence_transformers import SentenceTransformer
+
+# Sorgu tarafi prompt ayarlari (dokuman tarafi embed_index.py'de)
+MODELS = {
+    "magibu": ("magibu/embeddingmagibu-200m", "query", lambda t: t),
+    "e5":     ("ytu-ce-cosmos/turkish-e5-large", None, lambda t: f"query: {t}"),
+}
+
+ABSTAIN_MESSAGE = "Bu sorunun cevabi dokumanlarimda yer almamaktadir."
+
+
+def evaluate_model(key, model_name, prompt_name, prefix_fn, test_set, client, root):
+    model = SentenceTransformer(model_name, trust_remote_code=True)
+    col = client.get_collection(f"chunks_{key}")
+
+    def embed(text):
+        kw = {"prompt_name": prompt_name} if prompt_name else {}
+        return model.encode(prefix_fn(text), normalize_embeddings=True, **kw).tolist()
+
+    rows = []
+    for q in test_set:
+        res = col.query(query_embeddings=[embed(q["question"])], n_results=5)
+        top_ids = res["ids"][0]
+        top_sims = [1 - d for d in res["distances"][0]]  # cosine distance -> similarity
+        gt = q["ground_truth_chunk"]
+        hit_rank = top_ids.index(gt) + 1 if gt and gt in top_ids else None
+        rows.append({"id": q["id"], "label": q["label"],
+                     "top1_sim": top_sims[0], "hit_rank": hit_rank})
+    df = pd.DataFrame(rows)
+
+    pos = df[df.label == "positive"]["top1_sim"].values
+    neg = df[df.label == "negative"]["top1_sim"].values
+
+    # F1 sweep: esik gozle degil, veri-gudumlu secilir
+    best_t, best_f1, f1_curve = 0.0, 0.0, []
+    thresholds = np.linspace(0, 1, 200)
+    for t in thresholds:
+        tp, fp = int((pos >= t).sum()), int((neg >= t).sum())
+        fn = int((pos < t).sum())
+        prec = tp / (tp + fp) if tp + fp else 0.0
+        rec = tp / (tp + fn) if tp + fn else 0.0
+        f1 = 2 * prec * rec / (prec + rec) if prec + rec else 0.0
+        f1_curve.append(f1)
+        if f1 > best_f1:
+            best_f1, best_t = f1, t
+
+    tp, fp = int((pos >= best_t).sum()), int((neg >= best_t).sum())
+    fn, tn = int((pos < best_t).sum()), int((neg < best_t).sum())
+    pos_df = df[df.label == "positive"]
+    mrr = pos_df["hit_rank"].map(lambda r: 1 / r if pd.notna(r) else 0).mean()
+
+    # Grafik: skor dagilimi + F1 egrisi
+    fig, axes = plt.subplots(1, 2, figsize=(13, 4))
+    axes[0].hist(pos, bins=12, alpha=0.7, color="#2ecc71", label="Pozitif")
+    axes[0].hist(neg, bins=8, alpha=0.7, color="#e74c3c", label="Negatif")
+    axes[0].set_xlabel("Cosine Benzerlik"); axes[0].legend()
+    axes[0].set_title(f"{key} - Skor Dagilimi")
+    axes[1].plot(thresholds, f1_curve, color="#3498db")
+    axes[1].axvline(best_t, color="red", linestyle="--",
+                    label=f"t={best_t:.3f} (F1={best_f1:.3f})")
+    axes[1].set_xlabel("Threshold"); axes[1].set_ylabel("F1"); axes[1].legend()
+    axes[1].set_title(f"{key} - F1 Sweep")
+    plt.tight_layout()
+    plt.savefig(root / f"reports/threshold_{key}.png", dpi=120)
+    plt.close()
+
+    return {
+        "threshold": round(best_t, 3), "f1": round(best_f1, 3),
+        "precision": round(tp / (tp + fp), 3) if tp + fp else 0,
+        "recall": round(tp / (tp + fn), 3) if tp + fn else 0,
+        "tp": tp, "fp": fp, "fn": fn, "tn": tn,
+        "recall_at_1": round((pos_df.hit_rank == 1).sum() / len(pos_df), 3),
+        "recall_at_3": round((pos_df.hit_rank <= 3).sum() / len(pos_df), 3),
+        "recall_at_5": round((pos_df.hit_rank <= 5).sum() / len(pos_df), 3),
+        "mrr": round(mrr, 3),
+    }
+
+
+def run(config_path: str = "config.yaml") -> None:
+    root = Path(config_path).resolve().parent
+    with open(root / "data/test_set.jsonl", encoding="utf-8") as f:
+        test_set = [json.loads(l) for l in f]
+    client = chromadb.PersistentClient(path=str(root / "data/chroma"))
+    (root / "reports").mkdir(exist_ok=True)
+
+    results = {}
+    for key, (name, prompt, prefix) in MODELS.items():
+        print(f"\n=== {key} ===")
+        results[key] = evaluate_model(key, name, prompt, prefix,
+                                      test_set, client, root)
+        for m, v in results[key].items():
+            print(f"  {m}: {v}")
+
+    with open(root / "data/eval_results.json", "w", encoding="utf-8") as f:
+        json.dump(results, f, ensure_ascii=False, indent=2)
+    print("\n[OK] data/eval_results.json + reports/threshold_*.png")
+
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--config", default="config.yaml")
+    run(ap.parse_args().config)