Spaces:
Runtime error
Runtime error
| """ | |
| Tests du harnais d'evaluation. | |
| Ce qui est verifie ici, c'est que le harnais REFUSE de produire des chiffres | |
| faux : jeu non relu, corpus qui a bouge, recoupement lexical trop eleve. Un | |
| harnais qui accepte tout produit des nombres, pas des mesures. | |
| """ | |
| import json | |
| import numpy as np | |
| import pytest | |
| from scripts.eval.run_eval import ( | |
| _brute_force_rankings, | |
| _load_eval_set, | |
| _slice_and_renormalise, | |
| ) | |
| def _write_set(path, items, snapshot=None): | |
| path.write_text(json.dumps({ | |
| "version": 1, | |
| "corpus_snapshot": snapshot or {"laws": 1, "articles": 1, "alembic_head": "abc"}, | |
| "items": items, | |
| }, ensure_ascii=False), encoding="utf-8") | |
| return path | |
| def _item(**kwargs): | |
| base = { | |
| "question": "Quelle est la responsabilité des dirigeants ?", | |
| "expected_article_id": 1, | |
| "expected_law_id": 1, | |
| "expected_article_number": "161", | |
| "lexical_overlap": 0.1, | |
| "reviewed": True, | |
| "split": "dev", | |
| } | |
| base.update(kwargs) | |
| return base | |
| class TestGuards: | |
| def test_missing_file_is_explicit(self, tmp_path): | |
| with pytest.raises(SystemExit, match="absent"): | |
| _load_eval_set(tmp_path / "nexiste.pas", "dev", False, True) | |
| def test_unreviewed_items_are_refused(self, tmp_path): | |
| path = _write_set(tmp_path / "set.json", [_item(reviewed=False)]) | |
| with pytest.raises(SystemExit, match="non relus"): | |
| _load_eval_set(path, "dev", False, True) | |
| def test_unreviewed_can_be_forced_explicitly(self, tmp_path): | |
| path = _write_set(tmp_path / "set.json", [_item(reviewed=False)]) | |
| payload = _load_eval_set(path, "dev", True, True) | |
| assert len(payload["items"]) == 1 | |
| async def test_stale_corpus_snapshot_is_refused(self, tmp_path, db_session): | |
| """ | |
| Les identifiants d'article ne designent quelque chose que contre un etat | |
| de corpus donne. Sans ce controle, les chiffres deviennent faux en | |
| silence apres une reingestion. | |
| """ | |
| path = _write_set( | |
| tmp_path / "set.json", [_item()], | |
| snapshot={"laws": 999, "articles": 999999, "alembic_head": "perimee"}, | |
| ) | |
| with pytest.raises(SystemExit, match="corpus a change"): | |
| _load_eval_set(path, "dev", False, False) | |
| def test_split_selection(self, tmp_path): | |
| path = _write_set(tmp_path / "set.json", [ | |
| _item(split="dev", expected_article_id=1), | |
| _item(split="test", expected_article_id=2), | |
| ]) | |
| assert len(_load_eval_set(path, "dev", False, True)["items"]) == 1 | |
| assert len(_load_eval_set(path, "test", False, True)["items"]) == 1 | |
| assert len(_load_eval_set(path, "all", False, True)["items"]) == 2 | |
| class TestMatryoshkaSlicing: | |
| def test_slicing_renormalises(self): | |
| """ | |
| Le prefixe d'un vecteur unitaire n'est PAS unitaire : sans | |
| renormalisation, toute la comparaison cosinus est fausse. | |
| """ | |
| rng = np.random.default_rng(0) | |
| matrix = rng.normal(size=(5, 3072)).astype(np.float32) | |
| matrix /= np.linalg.norm(matrix, axis=1, keepdims=True) | |
| sliced = _slice_and_renormalise(matrix, 768) | |
| assert sliced.shape == (5, 768) | |
| assert np.allclose(np.linalg.norm(sliced, axis=1), 1.0, atol=1e-5) | |
| def test_zero_vector_does_not_divide_by_zero(self): | |
| matrix = np.zeros((2, 3072), dtype=np.float32) | |
| sliced = _slice_and_renormalise(matrix, 768) | |
| assert np.isfinite(sliced).all() | |
| class TestBruteForce: | |
| def test_ranks_by_cosine_similarity(self): | |
| article_ids = np.array([10, 20, 30]) | |
| corpus = np.array([ | |
| [1.0, 0.0], | |
| [0.0, 1.0], | |
| [0.7071, 0.7071], | |
| ], dtype=np.float32) | |
| queries = np.array([[1.0, 0.0]], dtype=np.float32) | |
| rankings = _brute_force_rankings(article_ids, corpus, queries, top_k=3) | |
| assert rankings == [[10, 30, 20]] | |
| def test_respects_top_k(self): | |
| article_ids = np.array([1, 2, 3, 4]) | |
| corpus = np.eye(4, dtype=np.float32) | |
| queries = np.array([[1.0, 0.0, 0.0, 0.0]], dtype=np.float32) | |
| assert len(_brute_force_rankings(article_ids, corpus, queries, top_k=2)[0]) == 2 | |