| """ |
| pusula_board.py — Sonuc dosyalarindan yayinlanabilir leaderboard uretir (metin icermez: yalnizca skorlar ve ID'ler). |
| |
| Siralama: mobil sorgularda mutlak basari. Komsu iki modelin farki VERI SETI DENGELI eslestirilmis bootstrap ile |
| test edilir; %95 guven araligi sifiri kapsiyorsa iki model ayni gruptadir (yapay siralama iddia edilmez). |
| """ |
| import numpy as np |
|
|
| from pusula_eval import robustness_row |
|
|
| LEXICAL_PREFIX = "bm25" |
|
|
|
|
| def balanced_paired_diff(a_by_ds, b_by_ds, n_boot=2000, seed=0): |
| """Her veri seti esit agirlikli: fark = mean_d(mean(a_d - b_d)). Bootstrap her sette ayri ornekler.""" |
| rng = np.random.default_rng(seed) |
| boots, point = np.zeros(n_boot), 0.0 |
| for a, b in zip(a_by_ds, b_by_ds): |
| d = np.asarray(a, float) - np.asarray(b, float) |
| point += d.mean() |
| boots += d[rng.integers(0, len(d), size=(n_boot, len(d)))].mean(axis=1) |
| k = len(a_by_ds) |
| return point / k, (float(np.percentile(boots / k, 2.5)), float(np.percentile(boots / k, 97.5))) |
|
|
|
|
| def _vec(pq, ds, pert, metric, qids): |
| return np.array([pq[ds][pert][metric][q] for q in qids]) |
|
|
|
|
| def build_board(results, datasets, changed, perts, metric="ndcg@10", rank_pert="mobile", |
| families=None, n_boot=2000, seed=0, holdout=()): |
| """results: load_results() ciktisi. datasets: {ad: {"queries": [qid,...], ...meta}}. |
| holdout: veri seti bazinda hesaplanip gosterilen ama ORTALAMAYA, SIRALAMAYA ve GRUPLARA girmeyen setler |
| (or. egitim verisiyle ortusen msmarco-tr).""" |
| families = families or {} |
| qids = {d: sorted(datasets[d]["queries"]) for d in datasets} |
| agg = [d for d in datasets if d not in set(holdout)] |
| assert agg, "En az bir veri seti siralamaya girmeli" |
| models = [] |
| for short, r in results.items(): |
| pq, spec = r["per_query"], r["spec"] |
| fam = families.get(short) or ("Sözcüksel" if short.startswith(LEXICAL_PREFIX) else "Diğer") |
| entry = {"short": short, "model_id": spec["model_id"], "params_m": spec.get("params_m") or None, |
| "revision": spec.get("revision"), "family": fam, "prompts": r.get("prompts"), |
| "dtype": r.get("dtype"), "by_dataset": {}} |
| for d in datasets: |
| clean_d = float(_vec(pq, d, "clean", metric, qids[d]).mean()) |
| row = {"clean": clean_d} |
| for p in perts: |
| rr = robustness_row(pq[d]["clean"], pq[d][p], changed[d][p], metric) |
| row[p] = {"score": float(rr["bozuk"]), |
| "retention": None if np.isnan(rr["koruma_değişen"]) else float(rr["koruma_değişen"]), |
| "changed_frac": float(rr["değişen_oran"]), |
| "diff_ci": [float(x) for x in rr["fark_değişen_ci"]] if rr["değişen_oran"] > 0 else None} |
| entry["by_dataset"][d] = row |
| entry["clean"] = float(np.mean([entry["by_dataset"][d]["clean"] for d in agg])) |
| entry["by_pert"] = {} |
| for p in perts: |
| rets = [entry["by_dataset"][d][p]["retention"] for d in agg if entry["by_dataset"][d][p]["retention"] is not None] |
| entry["by_pert"][p] = {"score": float(np.mean([entry["by_dataset"][d][p]["score"] for d in agg])), |
| "retention": float(np.mean(rets)) if rets else None} |
| vals = [v["retention"] for v in entry["by_pert"].values() if v["retention"] is not None] |
| entry["avg_retention"] = float(np.mean(vals)) if vals else None |
| entry["rank_score"] = entry["by_pert"][rank_pert]["score"] |
| models.append(entry) |
|
|
| models.sort(key=lambda m: -m["rank_score"]) |
| group = 1 |
| for i, m in enumerate(models): |
| m["rank"] = i + 1 |
| if i > 0: |
| prev = models[i - 1] |
| diff, ci = balanced_paired_diff( |
| [_vec(results[prev["short"]]["per_query"], d, rank_pert, metric, qids[d]) for d in agg], |
| [_vec(results[m["short"]]["per_query"], d, rank_pert, metric, qids[d]) for d in agg], |
| n_boot=n_boot, seed=seed) |
| prev["vs_next"] = {"diff": diff, "ci": list(ci), "significant": ci[0] > 0} |
| if ci[0] > 0: |
| group += 1 |
| m["group"] = group |
| clean_rank = {m["short"]: i + 1 for i, m in enumerate(sorted(models, key=lambda m: -m["clean"]))} |
| for m in models: |
| m["clean_rank"] = clean_rank[m["short"]] |
| return models |
|
|
|
|
| def findings(models, perts): |
| """Kahraman alanindaki cumleler icin veriden turetilen bulgular (sabit metin yok).""" |
| dense = [m for m in models if m["family"] != "Sözcüksel"] |
| out = {} |
| |
| scored = lambda p: [m for m in dense if m["by_pert"][p]["retention"] is not None] |
| if dense: |
| pairs = [(m, p) for p in perts for m in scored(p)] |
| if pairs: |
| worst = min(pairs, key=lambda x: x[0]["by_pert"][x[1]]["retention"]) |
| out["worst_drop"] = {"model": worst[0]["short"], "pert": worst[1], |
| "retention": worst[0]["by_pert"][worst[1]]["retention"]} |
| best_clean = min(dense, key=lambda m: m["clean_rank"]) |
| out["clean_winner"] = {"model": best_clean["short"], "clean_rank": best_clean["clean_rank"], "mobile_rank": best_clean["rank"]} |
| out["mobile_winner"] = {"model": models[0]["short"], "clean_rank": models[0]["clean_rank"], "group": models[0]["group"], |
| "tied_with": [m["short"] for m in models[1:] if m["group"] == models[0]["group"]]} |
| out["most_fragile"] = {p: min(scored(p), key=lambda m: m["by_pert"][p]["retention"])["short"] for p in perts if scored(p)} |
| out["most_robust"] = {p: max(scored(p), key=lambda m: m["by_pert"][p]["retention"])["short"] for p in perts if scored(p)} |
| lex = {m["short"]: m for m in models if m["family"] == "Sözcüksel"} |
| if "bm25" in lex and "bm25-fold" in lex and "ascii_full" in perts: |
| out["fold_effect"] = {"before": lex["bm25"]["by_pert"]["ascii_full"]["retention"], |
| "after": lex["bm25-fold"]["by_pert"]["ascii_full"]["retention"]} |
| return out |
|
|
|
|
| def export_payload(models, perts, labels, descriptions, datasets_meta, fingerprint, dataset_repo, |
| example_text, example_outputs, failed, metric="ndcg@10", failed_notes=None, holdout=None): |
| """Space ve dataset reposuna gidecek tek JSON. Benchmark metni ICERMEZ (ornek cumle bize ait). |
| holdout: {veri_seti: neden} — siralamaya girmeyen setler; datasets_meta'ya in_ranking/holdout_reason eklenir.""" |
| failed_notes, holdout = failed_notes or {}, holdout or {} |
| datasets_meta = [{**d, "in_ranking": d["name"] not in holdout, "holdout_reason": holdout.get(d["name"])} |
| for d in datasets_meta] |
| fails = [] |
| for f in failed: |
| reason = failed_notes.get(f["short"]) or (f.get("error") or "bilinmeyen hata").splitlines()[0][:160] |
| fails.append({"short": f["short"], "model_id": f["model_id"], "reason": reason}) |
| return { |
| "meta": {"fingerprint": fingerprint, "metric": metric, "dataset_repo": dataset_repo, |
| "perturbations": [{"name": p, "label": labels.get(p, p), "description": descriptions.get(p, "")} for p in perts], |
| "datasets": datasets_meta, "failed": fails, |
| "example": {"text": example_text, "outputs": example_outputs}}, |
| "models": models, |
| "findings": findings(models, perts), |
| } |
|
|
