llm-memory-editability / docs /development-artifacts /hebbian-future-v1 /source /tests /test_hebbian_future.py
| import numpy as np | |
| from llm_memory_editability.hebbian_future import ( | |
| CONFIG, | |
| chain_world, | |
| read, | |
| ridge_predict, | |
| training_arrays, | |
| ) | |
| def test_balanced_worlds_and_heldout_composition(): | |
| cfg = read(CONFIG)["chains"] | |
| for world in cfg["worlds"]: | |
| previous = None | |
| for rho in cfg["correlations"]: | |
| w = chain_world(world, rho, cfg) | |
| for split in (w["train_people"], ~w["train_people"]): | |
| assert np.mean(w["home_y"][split] == w["composite_y"][split]) == rho | |
| assert len(set(np.bincount(w["home"][split], minlength=4))) == 1 | |
| assert ( | |
| len(set(np.bincount(w["composite_y"][split] - w["city_start"], minlength=4))) | |
| == 1 | |
| ) | |
| assert np.all( | |
| w["home_y"][w["common_conflict"]] != w["composite_y"][w["common_conflict"]] | |
| ) | |
| x, y = training_arrays(w) | |
| assert len(x) == 252 and len(y) == 252 | |
| train_composites = {tuple(row) for row in x[x[:, 2] == 6]} | |
| assert not train_composites.intersection( | |
| map(tuple, w["composite_x"][~w["train_people"]]) | |
| ) | |
| if previous is not None: | |
| for key in ( | |
| "member_x", | |
| "member_y", | |
| "root_x", | |
| "root_y", | |
| "composite_x", | |
| "composite_y", | |
| "train_people", | |
| "common_conflict", | |
| ): | |
| np.testing.assert_array_equal(w[key], previous[key]) | |
| np.testing.assert_array_equal(x, training_arrays(previous)[0]) | |
| previous = w | |
| def test_ridge_uses_only_development_statistics(): | |
| train = np.array([[0.0, 1.0], [1.0, 1.0], [2.0, 1.0], [3.0, 1.0]]) | |
| labels = np.array([0.0, 0.0, 1.0, 1.0]) | |
| test = np.array([[1.5, 1.0], [100.0, 1.0]]) | |
| together = ridge_predict(train, labels, test) | |
| separate = np.r_[ridge_predict(train, labels, test[:1]), ridge_predict(train, labels, test[1:])] | |
| np.testing.assert_allclose(together, separate) | |
| def test_two_step_uses_generated_membership(): | |
| import importlib.util | |
| import torch | |
| spec = importlib.util.spec_from_file_location("future_runner", "scripts/run_hebbian_future.py") | |
| runner = importlib.util.module_from_spec(spec) | |
| spec.loader.exec_module(runner) | |
| w = chain_world(300, 0.5, read(CONFIG)["chains"]) | |
| class WrongMembership(torch.nn.Module): | |
| def forward(self, tokens): | |
| logits = torch.zeros(len(tokens), tokens.shape[1], w["vocab"]) | |
| if tokens.shape[1] == 5: | |
| pred = torch.full((len(tokens),), 7) | |
| else: | |
| pred = torch.where( | |
| tokens[:, 2] == 3, | |
| w["group_start"] + 1, | |
| w["city_start"] + (tokens[:, 1] - w["group_start"]) % 4, | |
| ) | |
| logits[torch.arange(len(tokens)), -1, pred] = 10 | |
| return logits | |
| result = runner.evaluate_chain(WrongMembership(), w, "cpu") | |
| actual = np.array(result["rows"]["two_step"]["prediction"]) | |
| assert np.all(actual == w["city_start"] + 1) | |
| assert not np.array_equal(actual, w["composite_y"]) | |