| |
| """Lock QH-024 train/eval blocks and hyperparameters before formal runs.""" |
|
|
| from __future__ import annotations |
|
|
| import hashlib |
| import json |
| from datetime import datetime, timezone |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
|
|
|
|
| PROJECT = Path(__file__).resolve().parents[2] |
| ARTIFACTS = PROJECT / "artifacts" |
| STAGE_A = ARTIFACTS / "wikitext_stage_a" |
| TRAIN = PROJECT / "datasets/processed/wikitext2_qwen38/train-tokens.pt" |
| VALIDATION = PROJECT / "datasets/processed/wikitext2_qwen38/validation-tokens.pt" |
| PARENT_LOCK = ARTIFACTS / "wikitext-replacement-screen-lock.json" |
| SCREEN = ARTIFACTS / "wikitext-svd-value-replacement-screen.json" |
| OUTPUT = ARTIFACTS / "wikitext-qh024-distillation-lock.json" |
| SEED = 20260838 |
| SEQUENCE_LENGTH = 256 |
| TRAIN_COUNT = 512 |
| EVAL_COUNT = 32 |
|
|
|
|
| def sha256(path: Path) -> str: |
| digest = hashlib.sha256() |
| with path.open("rb") as handle: |
| for chunk in iter(lambda: handle.read(8 << 20), b""): |
| digest.update(chunk) |
| return digest.hexdigest() |
|
|
|
|
| def prior_train_indices() -> set[int]: |
| paths = [ |
| STAGE_A / "wikitext-qh010-s20260829-n256-l256.json", |
| STAGE_A / "wikitext-cc010-s20260829-n256-l256.json", |
| STAGE_A / "wikitext-qh022-unused-complement-s20260829-n256-l256.json", |
| STAGE_A / "wikitext-cc022-unused-complement-s20260829-n256-l256.json", |
| ] |
| values: set[int] = set() |
| for path in paths: |
| if path.exists(): |
| payload = json.loads(path.read_text(encoding="utf-8")) |
| values.update(int(index) for index in payload["dataset"]["train_indices"]) |
| return values |
|
|
|
|
| def main() -> None: |
| parent = json.loads(PARENT_LOCK.read_text(encoding="utf-8")) |
| if parent["status"] != "locked_before_any_replacement_performance_result": |
| raise RuntimeError("invalid parent replacement lock") |
| if not SCREEN.exists(): |
| raise RuntimeError("SVD screen record is required before architecture locking") |
| eval_indices = [int(value) for value in parent["indices"][32:64]] |
| if len(eval_indices) != EVAL_COUNT: |
| raise RuntimeError("parent lock does not contain the reserved holdout") |
| screen_payload = json.loads(SCREEN.read_text(encoding="utf-8")) |
| screened_indices = { |
| int(row["block_index"]) for row in screen_payload["base"]["rows"] |
| } |
| if screened_indices & set(eval_indices): |
| raise RuntimeError("QH-024 holdout overlaps the SVD selection blocks") |
|
|
| train_tokens = torch.load(TRAIN, map_location="cpu", weights_only=True) |
| complete_train = len(train_tokens) // SEQUENCE_LENGTH |
| excluded_train = prior_train_indices() |
| available = np.array(sorted(set(range(complete_train)) - excluded_train), dtype=np.int64) |
| train_indices = np.random.default_rng(SEED).permutation(available)[:TRAIN_COUNT].tolist() |
| if len(train_indices) != TRAIN_COUNT or set(train_indices) & excluded_train: |
| raise RuntimeError("failed to create disjoint QH-024 training blocks") |
|
|
| payload = { |
| "status": "locked_before_qh024_qh_cc_base_performance_results", |
| "candidate": "QH-024 true v_proj replacement", |
| "architecture": { |
| "layer_index": 7, |
| "target": "v_proj", |
| "svd_rank": 512, |
| "qubits": 10, |
| "depth": 2, |
| "quantum_trainable_parameters": 112701, |
| "original_projection_parameters": 5242880, |
| "deployed_replacement_parameters": 3258429, |
| "net_model_parameter_reduction": 1984451, |
| }, |
| "selection_basis": { |
| "screen_record": str(SCREEN), |
| "screen_sha256": sha256(SCREEN), |
| "selected_candidate": "svd-v-l7-r512", |
| "selection_rule": ( |
| "largest tested parameter reduction among candidates with development-screen " |
| "mean NLL increase <= 0.005" |
| ), |
| }, |
| "dataset": "Salesforce/wikitext wikitext-2-raw-v1", |
| "source_revision": "b08601e04326c79dfdd32d625aee71d232d685c3", |
| "license": "CC BY-SA 4.0", |
| "sequence_length": SEQUENCE_LENGTH, |
| "seed": SEED, |
| "train_indices": train_indices, |
| "train_count": TRAIN_COUNT, |
| "excluded_prior_train_indices_count": len(excluded_train), |
| "train_overlap_with_excluded": len(set(train_indices) & excluded_train), |
| "eval_indices": eval_indices, |
| "eval_count": EVAL_COUNT, |
| "eval_overlap_with_svd_selection_blocks": len(screened_indices & set(eval_indices)), |
| "parent_lock": str(PARENT_LOCK), |
| "parent_lock_sha256": sha256(PARENT_LOCK), |
| "train_tokens_sha256": sha256(TRAIN), |
| "validation_tokens_sha256": sha256(VALIDATION), |
| "training_protocol": { |
| "objective": "relative MSE distillation of the deleted dense v_proj output", |
| "optimizer": "AdamW", |
| "learning_rate": 0.001, |
| "weight_decay": 0.0, |
| "gradient_clip_norm": 1.0, |
| "epochs": 1, |
| "all_original_model_parameters_frozen": True, |
| "teacher_dense_weight_is_training_only_nonpersistent_buffer": True, |
| }, |
| "evaluation_protocol": { |
| "primary_metric": "mean token NLL", |
| "compression_noninferiority_margin_nll": 0.005, |
| "paired_bootstrap_samples": 20000, |
| "required_controls": ["frozen_dense_base", "svd512_zero_branch", "CC-024"], |
| }, |
| "test_rows_inspected_or_used_for_fitness": 0, |
| "created_at": datetime.now(timezone.utc).isoformat(), |
| } |
| OUTPUT.write_text(json.dumps(payload, indent=2), encoding="utf-8") |
| print(json.dumps({ |
| "status": payload["status"], |
| "train_count": TRAIN_COUNT, |
| "eval_count": EVAL_COUNT, |
| "train_overlap": payload["train_overlap_with_excluded"], |
| "eval_overlap": payload["eval_overlap_with_svd_selection_blocks"], |
| "record": str(OUTPUT), |
| "sha256": sha256(OUTPUT), |
| }, indent=2)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
