| import json |
|
|
| import pytest |
|
|
| from gemma_rlcd.cached_backend import CachedMLXBackend |
| from gemma_rlcd.catalog_backend import CatalogMLXBackend |
| from gemma_rlcd.core import ScoringRequest, State |
|
|
|
|
| def test_catalog_retains_complete_descriptions_and_uses_short_selectors(monkeypatch): |
| backend = CatalogMLXBackend.__new__(CatalogMLXBackend) |
| backend.branch_batch_size = 8 |
| requests = [ |
| ScoringRequest( |
| "original", |
| ("A", "B"), |
| "Question\nwith detail", |
| (("yes", "First\nsecond line"), ("no", "No")), |
| ) |
| ] |
| captured = {} |
|
|
| def prepare(self, state, selectors): |
| captured["catalog"] = json.loads(self._catalog) |
| captured["selectors"] = selectors |
| return "prepared" |
|
|
| monkeypatch.setattr(CachedMLXBackend, "prepare", prepare) |
| assert backend.prepare(State(text="evidence"), requests) == "prepared" |
| assert captured["catalog"][0]["question"] == "Question\nwith detail" |
| assert captured["catalog"][0]["options"][0]["meaning"] == "First\nsecond line" |
| assert captured["selectors"][0].prompt == "Evaluate field 0. Answer code:" |
| assert backend._catalog is None |
|
|
|
|
| def test_catalog_rejects_overflow_instead_of_serializing_fields(): |
| backend = CatalogMLXBackend.__new__(CatalogMLXBackend) |
| backend.branch_batch_size = 1 |
| with pytest.raises(ValueError, match="one branch batch"): |
| backend.prepare(State(text="evidence"), [ScoringRequest("x", ("A", "B"))] * 2) |
|
|
