File size: 11,678 Bytes
53a1159 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 | """Contracts affecting truth, end-to-end text, scoring and paired depth initialization."""
import copy
import json
import numpy as np
import pytest
import torch
from torch.nn import functional as F
from llm_memory_editability.depth_step import (
BOS,
EOS,
EVALUATION_SPLITS,
TRAIN_SPLITS,
_validate_output,
audit_world,
build_world,
construct,
data_digest,
evaluate,
generate_rows,
pack_rows,
prompt_rows,
run_name,
truth_path_details,
)
from llm_memory_editability.grok_depth import EpochStream, SmallGPT
from llm_memory_editability.grok_loop_data import build_world as old_world
@pytest.fixture(scope="module")
def spec():
return {
"world_seed": 751011,
"entities": 64,
"relations": 4,
"degree": 4,
"phi": 4.0,
"id_fraction": 0.75,
"id_test_fraction": 0.2,
"evaluation_size": 512,
"width": 32,
"heads": 4,
"layers": 2,
"repeats": 1,
"dropout": 0.0,
"initialization": 752011,
"stream_seed": 753011,
"batch_size": 128,
"steps": 64000,
}
@pytest.fixture(scope="module")
def world(spec):
return build_world(spec)
def test_three_lengths_share_exact_graph_id_mask_and_full_reserved_pools(spec, world):
report = audit_world(world)
assert report["counts"] == {
"atomic": 256,
"train_2": 444,
"familiar_2": 124,
"strict_2": 64,
"train_3": 768,
"familiar_3": 329,
"strict_3": 59,
"train_4": 768,
"familiar_4": 1025,
"strict_4": 53,
}
for hops in (2, 3, 4):
source = old_world(dict(spec, hops=hops))
np.testing.assert_array_equal(world["atomic"], source["atomic"] + 1)
np.testing.assert_array_equal(world[f"familiar_{hops}"], source["test_full_composite"] + 1)
np.testing.assert_array_equal(world[f"strict_{hops}"], source["ood_composite"] + 1)
source_ids = set(map(tuple, source["id_atomic"]))
expected = [tuple(row) in source_ids for row in source["atomic"]]
assert world["metadata"]["id_mask"] == expected
assert (
world["metadata"]["source_worlds"][str(hops)]["dataset_sha256"]
== source["metadata"]["dataset_sha256"]
)
assert world["metadata"]["source_token_shift"] == 1
assert world["metadata"]["source_worlds"]["2"]["train_composite_capped"]
def test_truth_edges_and_strict_facts_never_used_at_any_training_length(world):
trained = set()
lookup = {(int(h), int(r)): int(t) for h, r, t in world["atomic"]}
for hops in (2, 3, 4):
rows = world[f"train_{hops}"]
nodes, edges = truth_path_details(world, rows)
assert nodes.shape == (len(rows), hops + 1)
assert edges.shape == (len(rows), hops)
trained.update(edges.ravel())
for i, row in enumerate(rows):
head = int(row[0])
for j, relation in enumerate(row[1:-1]):
head = lookup[head, int(relation)]
assert head == nodes[i, j + 1]
assert head == row[-1]
for hops in (2, 3, 4):
strict = truth_path_details(world, world[f"strict_{hops}"])[1]
assert not set(strict.ravel()) & trained
@pytest.mark.parametrize("change", ["truth", "id_mask", "overlap", "duplicate"])
def test_invalid_data_contracts_are_rejected(world, change):
damaged = copy.deepcopy(world)
if change == "truth":
tail = damaged["familiar_3"][0, -1]
damaged["familiar_3"][0, -1] = 3 + (tail - 3 + 1) % 64
elif change == "id_mask":
edge = truth_path_details(damaged, damaged["train_2"][:1])[1][0, 0]
damaged["metadata"]["id_mask"][edge] = False
elif change == "overlap":
damaged["familiar_2"][0] = damaged["train_2"][0]
else:
damaged["train_4"][1] = damaged["train_4"][0]
with pytest.raises(ValueError):
audit_world(damaged)
def test_frozen_digest_enforced_and_architecture_does_not_change_world(spec, world):
digest = data_digest(world)
for layers, repeats in ((1, 4), (6, 1)):
actual = build_world(dict(spec, layers=layers, repeats=repeats, frozen_data_sha256=digest))
assert data_digest(actual) == digest
with pytest.raises(ValueError, match="frozen"):
build_world(dict(spec, frozen_data_sha256="invalid"))
def test_compact_prompt_has_separator_full_token_ce_and_no_intermediate_labels(world):
separator = world["metadata"]["separator_token"]
assert separator == 71
assert world["metadata"]["vocab_size"] == 72
for split in TRAIN_SPLITS:
rows = world[split][:3]
prompts = prompt_rows(rows, separator)
assert np.all(prompts[:, 0] == BOS)
assert np.all(prompts[:, -1] == separator)
np.testing.assert_array_equal(prompts[:, 1:-1], rows[:, :-1])
tokens, labels = pack_rows(rows, separator=separator)
length = rows.shape[1] + 2
expected = np.c_[
rows[:, :-1], np.full(len(rows), separator), rows[:, -1], np.full(len(rows), EOS)
]
np.testing.assert_array_equal(labels[:, :length], expected)
assert np.all(labels[:, length:] == -100)
assert np.all(tokens[:, length:] == 0)
assert not (labels == BOS).any()
assert len(pack_rows(world["train_4"][:1])[0][0]) == 8
assert sum(int((pack_rows(world[key][:1])[1] >= 0).sum()) for key in TRAIN_SPLITS) * 32 == 832
with pytest.raises(ValueError):
pack_rows(world["train_4"][:1], sequence=7)
def test_paired_initialization_all_architectures_and_effective_depth(spec):
torch.set_num_threads(1)
architectures = [(layers, 1) for layers in (1, 2, 3, 4, 6)] + [(1, r) for r in (2, 3, 4, 6)]
models = {
(layers, r): construct(dict(spec, layers=layers, repeats=r), "cpu")
for layers, r in architectures
}
reference = models[1, 1]
for (layers, repeats), model in models.items():
assert len(model.blocks) == layers and model.effective_depth == layers * repeats
assert all(p.requires_grad for p in model.parameters())
for name in ("token.weight", "position.weight", "ln_final.weight", "ln_final.bias"):
torch.testing.assert_close(
model.state_dict()[name], reference.state_dict()[name], rtol=0, atol=0
)
for depth in (2, 3, 4, 6):
ordinary, loop = models[depth, 1], models[1, depth]
for key, value in ordinary.blocks[0].state_dict().items():
torch.testing.assert_close(value, loop.blocks[0].state_dict()[key], rtol=0, atol=0)
assert not torch.equal(
models[2, 1].blocks[0].mlp.up.weight, models[2, 1].blocks[1].mlp.up.weight
)
assert len({run_name(dict(spec, layers=layers, repeats=r)) for layers, r in architectures}) == 9
def test_standard_full_forward_causal_and_all_parameters_train(spec, world):
torch.set_num_threads(1)
model = construct(spec, "cpu").eval()
ordinary = SmallGPT(model.config, dropout=0.0).eval()
ordinary.load_state_dict(model.state_dict())
tokens = torch.as_tensor(pack_rows(world["train_4"][:2])[0])
logits = model(tokens)
torch.testing.assert_close(logits, ordinary(tokens), rtol=0, atol=0)
assert logits.shape == (2, 8, 72)
changed = tokens.clone()
changed[:, 5:] = 3
torch.testing.assert_close(model(changed)[:, :5], logits[:, :5], rtol=0, atol=0)
labels = torch.as_tensor(pack_rows(world["train_4"][:2])[1])
F.cross_entropy(logits.flatten(0, 1), labels.flatten(), ignore_index=-100).backward()
for param in (
model.token.weight,
model.position.weight,
model.blocks[0].attention.qkv.weight,
model.blocks[0].mlp.up.weight,
model.blocks[0].mlp.down.weight,
):
assert (
param.grad is not None
and torch.isfinite(param.grad).all()
and param.grad.abs().sum() > 0
)
class TruthOracle(torch.nn.Module):
def __init__(self, world, wrong_eos=False):
super().__init__()
self.lookup = {(int(h), int(r)): int(t) for h, r, t in world["atomic"]}
self.separator = world["metadata"]["separator_token"]
self.vocab = world["metadata"]["vocab_size"]
self.wrong_eos = wrong_eos
def forward(self, tokens):
logits = torch.full((*tokens.shape, self.vocab), -20.0, device=tokens.device)
for i, row in enumerate(tokens.tolist()):
if row[-1] == self.separator:
current = row[1]
for relation in row[2:-1]:
current = self.lookup.get((current, relation), 3)
result = current
else:
result = 0 if self.wrong_eos else EOS
logits[i, -1, result] = 20.0
return logits
def test_free_tail_eos_coverage_and_autonomous_truth_scoring(world):
oracle = TruthOracle(world)
metrics, predictions = evaluate(oracle, world, "cpu")
assert oracle.training
for name in EVALUATION_SPLITS:
assert metrics[name]["accuracy"] == metrics[name]["answer_accuracy"] == 1.0
assert predictions[name + "_correct"].all()
if name != "atomic":
assert metrics[name]["atomic_correct_coverage"] == 1.0
assert metrics[name]["conditional_accuracy"] == 1.0
assert metrics[name]["autonomous_two_calls"] == 1.0
assert metrics[name]["autonomous_path_accuracy"] == 1.0
metrics, _ = evaluate(TruthOracle(world, wrong_eos=True), world, "cpu")
assert metrics["atomic"]["answer_accuracy"] == 1.0
assert metrics["atomic"]["accuracy"] == 0.0
assert metrics["familiar_4"]["conditional_accuracy"] is None
assert metrics["familiar_4"]["autonomous_two_calls"] == 0.0
def test_empty_pools_never_report_nan(world):
empty = dict(world)
empty["strict_4"] = np.empty((0, 6), dtype=np.int64)
metrics, predictions = evaluate(TruthOracle(world), empty, "cpu")
assert metrics["strict_4"]["n"] == 0
assert metrics["strict_4"]["accuracy"] is None
assert metrics["strict_4"]["atomic_correct_coverage"] is None
assert predictions["strict_4_autonomous_generated"].shape == (0, 4, 2)
json.dumps(metrics, allow_nan=False)
task, prediction = generate_rows(TruthOracle(world), empty["strict_4"], "cpu")
assert task["answer_nll"] is None and prediction["generated"].shape == (0, 2)
def test_epoch_stream_batches_and_saved_continuation_match_across_architectures(world, spec):
for i, name in enumerate(TRAIN_SPLITS):
first = EpochStream(len(world[name]), spec["stream_seed"] + i)
second = EpochStream(len(world[name]), spec["stream_seed"] + i)
np.testing.assert_array_equal(
first.take(2048).reshape(64, 32), second.take(2048).reshape(64, 32)
)
restored = EpochStream(len(world[name]), 0)
restored.load_state_dict(first.state_dict())
np.testing.assert_array_equal(first.take(32), restored.take(32))
@pytest.mark.parametrize(
"artifact",
[
"latest.pt",
"complete.json",
"world.npz",
"spec.json",
"predictions-000000.npz",
"exposures.npz",
],
)
def test_attempt_artifacts_never_overwritten_but_scheduler_files_allowed(tmp_path, artifact):
(tmp_path / "input-spec.json").write_text("{}")
(tmp_path / "train-process.log").write_text("scheduler")
_validate_output(tmp_path)
(tmp_path / artifact).write_bytes(b"existing experiment")
with pytest.raises(FileExistsError):
_validate_output(tmp_path)
assert (tmp_path / artifact).read_bytes() == b"existing experiment"
|