ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints (part 4)
b9301f1 verified | """Speech-lock repairs: no unigram broadcast, soup not re-ingested.""" | |
| from __future__ import annotations | |
| import inspect | |
| import unittest | |
| import torch | |
| from src.atom_native import AtomNativeModel, AtomSurfaceHead | |
| from src.io.atomizer import Atomizer | |
| from src.speech_lock import format_dialogue_prompt, speech_ok | |
| class SpeechLockUnitTests(unittest.TestCase): | |
| def test_dialogue_wrap(self) -> None: | |
| self.assertEqual( | |
| format_dialogue_prompt("Bonjour"), | |
| "Utilisateur: Bonjour\nAssistant:", | |
| ) | |
| self.assertIn("Assistant:", format_dialogue_prompt("Utilisateur: Hi\nAssistant:")) | |
| def test_speech_ok_rejects_attractor_soup(self) -> None: | |
| self.assertFalse(speech_ok(b"")) | |
| self.assertFalse(speech_ok(b"auiseateu : aernun e")) | |
| self.assertFalse(speech_ok(b"asssuatiur:i derouin")) | |
| self.assertTrue(speech_ok(b"Salut")) | |
| self.assertTrue(speech_ok(b"Je suis la.")) | |
| def test_payload_copy_is_position_aligned_not_broadcast(self) -> None: | |
| torch.manual_seed(0) | |
| head = AtomSurfaceHead(d_model=8, max_payload_bytes=8, n_modes=8) | |
| a_only = torch.ones(16) | |
| a_hat = a_only / a_only.norm().clamp_min(1e-8) | |
| byte_logits, _ = head.payload_produce( | |
| [b"abc"], | |
| a_hat, | |
| a_only, | |
| atom_energies=[1.0], | |
| ) | |
| aligned = "No unigram broadcast" in inspect.getsource(head.payload_produce) | |
| if not aligned: | |
| self.skipTest("run python3 scripts/apply_speech_lock.py") | |
| self.assertGreater( | |
| float(byte_logits[0, ord("a")]), | |
| float(byte_logits[3, ord("a")]), | |
| ) | |
| class SpeechGenerateTests(unittest.TestCase): | |
| def test_generate_does_not_commit_soup(self) -> None: | |
| sig = inspect.signature(AtomNativeModel.generate_packets) | |
| if "speech_gate" not in sig.parameters: | |
| self.skipTest("run python3 scripts/apply_speech_lock.py") | |
| model = AtomNativeModel( | |
| d_model=8, | |
| n_modes=8, | |
| n_atoms_max=32, | |
| max_payload_bytes=8, | |
| atomizer=Atomizer(max_span_bytes=8), | |
| ) | |
| model.eval() | |
| raw = model.generate_packets( | |
| "Bonjour", | |
| max_packets=4, | |
| deterministic=True, | |
| payload_copy=False, | |
| dialogue_wrap=True, | |
| speech_gate=True, | |
| ) | |
| self.assertNotIn(b"auiseateu", raw) | |
| if __name__ == "__main__": | |
| unittest.main() | |