{ "schema_version": 1, "protocol_id": "rosetta-published-coco-baseline-v1", "status": "pre_run_protocol", "recorded_at_utc": "2026-10-10T23:23:49.452212+00:00", "objective": "Reproduce one complete published COCO zero-pair DINOv2 B/14 mean \u00d7 MPNet configuration over all five original seeds, including full retrieval and classification evaluation. This is phase 1 of the approved replication-and-controls plan.", "upstream_commit": "fdfad84ea488f29702a61a455be71622c4ff5d49", "upstream_url": "https://github.com/dominik-schnaus/unpaired-rosetta/tree/fdfad84ea488f29702a61a455be71622c4ff5d49", "seeds": [ 734796314, 576165995, 2197670066, 839703249, 2584932063 ], "selection_filter": "method == 'ours' and dataset == 'coco_train2014' and model_x == 'dinov2_vit-b14@224_mean' and model_y == 'mpnet'", "configuration_except_seed": { "dataset": "coco_train2014", "model_x": "dinov2_vit-b14@224_mean", "model_y": "mpnet", "method": "ours", "modality_x": "vision", "modality_y": "language", "dataset_y": null, "num_pairs": 0, "pairs_dataset": null, "validation": "coco_val2014", "holdout": 0, "train_size": null, "validation_size": null, "classification": [ "CIFAR-10", "CIFAR-100", "ImageNet-100" ], "labels": null }, "algorithm_defaults": { "num_clusters": 30, "num_restarts": 30, "batch_size": 10000, "num_iterations": 100 }, "execution": { "workers": 4, "threads": 4, "environment": "default", "seeds_run_sequentially": true, "algorithm_device": "cpu", "required_command": "pixi run --frozen -e default python -m experiments.unpaired run --workers 4 --threads 4 --only FILTER", "source_changes_allowed": false }, "expected_python": "3.14.7", "expected_versions": { "torch": "2.14.0+cu130", "torchvision": "0.29.0+cu130", "numpy": "2.5.3", "scipy": "1.18.0", "scikit-learn": "1.9.1", "pylibmgm": "1.1.2", "pot": "0.9.7.post1", "transformers": "5.12.0", "timm": "1.0.29", "scotplus": "1.0.2" }, "expected_activation": { "MKL_CBWR": "AVX2,STRICT", "ATEN_CPU_CAPABILITY": "avx2" }, "expected_shapes": { "embeddings/coco_train2014/vision/dinov2_vit-b14@224_mean.pt": [ 82783, 768 ], "embeddings/coco_train2014/language/mpnet.pt": [ 82783, 7, 768 ], "embeddings/coco_val2014/vision/dinov2_vit-b14@224_mean.pt": [ 40504, 768 ], "embeddings/coco_val2014/language/mpnet.pt": [ 40504, 7, 768 ], "embeddings/CIFAR-10/vision/dinov2_vit-b14@224_mean.pt": [ 50000, 768 ], "embeddings/CIFAR-10/language_b953684b7ed3a62cbea9fe4e30935ee2/mpnet.pt": [ 10, 18, 768 ], "embeddings/CIFAR-10/labels.pt": [ 50000 ], "embeddings/CIFAR-100/vision/dinov2_vit-b14@224_mean.pt": [ 50000, 768 ], "embeddings/CIFAR-100/language_b953684b7ed3a62cbea9fe4e30935ee2/mpnet.pt": [ 100, 18, 768 ], "embeddings/CIFAR-100/labels.pt": [ 50000 ], "embeddings/ImageNet-100/vision/dinov2_vit-b14@224_mean.pt": [ 5000, 768 ], "embeddings/ImageNet-100/language/mpnet.pt": [ 100, 1, 768 ], "embeddings/ImageNet-100/labels.pt": [ 5000 ] }, "expected_dtypes": { "embeddings/coco_train2014/vision/dinov2_vit-b14@224_mean.pt": "torch.bfloat16", "embeddings/coco_train2014/language/mpnet.pt": "torch.bfloat16", "embeddings/coco_val2014/vision/dinov2_vit-b14@224_mean.pt": "torch.bfloat16", "embeddings/coco_val2014/language/mpnet.pt": "torch.bfloat16", "embeddings/CIFAR-10/vision/dinov2_vit-b14@224_mean.pt": "torch.bfloat16", "embeddings/CIFAR-10/language_b953684b7ed3a62cbea9fe4e30935ee2/mpnet.pt": "torch.bfloat16", "embeddings/CIFAR-10/labels.pt": "torch.int64", "embeddings/CIFAR-100/vision/dinov2_vit-b14@224_mean.pt": "torch.bfloat16", "embeddings/CIFAR-100/language_b953684b7ed3a62cbea9fe4e30935ee2/mpnet.pt": "torch.bfloat16", "embeddings/CIFAR-100/labels.pt": "torch.int64", "embeddings/ImageNet-100/vision/dinov2_vit-b14@224_mean.pt": "torch.bfloat16", "embeddings/ImageNet-100/language/mpnet.pt": "torch.bfloat16", "embeddings/ImageNet-100/labels.pt": "torch.int64" }, "sample_counts": { "train_corpus": 82783, "fit_x": 41391, "fit_y": 41392, "validation": 40504, "CIFAR-10": 50000, "CIFAR-100": 50000, "ImageNet-100": 5000 }, "required_metrics": [ "foscttm", "mean_rank", "num_validation_samples", "CIFAR-10_top1", "CIFAR-10_top5", "CIFAR-100_top1", "CIFAR-100_top5", "ImageNet-100_top1", "ImageNet-100_top5", "fit_and_evaluation_seconds" ], "input_preprocessing": "Use original load_rows/reduce_texts without edits: normalize each stored caption, average non-NaN captions, normalize, all in stored bfloat16 precision, then convert to float32. Preserve exact disjoint half split and the entire validation set.", "comparison": { "reference_file": "reference.json", "all_five_seeds_required": true, "aggregation": "Arithmetic mean and sample standard deviation ddof=1; no seed omission or best-seed selection.", "full_precision_foscttm": { "mean": 0.042989911199083145, "sample_sd": 0.0480892962510061, "specific_seed": 734796314, "specific_seed_value": 0.016328599083198735 }, "strict_numerical_agreement_absolute_tolerance": 1e-10, "strict_tolerance_attribution": "Our prespecified numerical comparison flag, not an author-provided tolerance or statistical test. Apply separately to every available full-precision retrieval reference.", "paper_display_agreement": { "foscttm_mean_and_sample_sd_decimal_places": 2, "classification_mean_and_sample_sd_percent_decimal_places": 0, "report_each_metric_separately": true }, "unavailable_references": "No complete five-seed reference list or full-precision classification aggregates located. Aggregate agreement cannot prove exact reproduction of all individual seeds.", "interpretation": "Report numerical differences and rounded agreement separately. A mismatch triggers implementation/environment/input investigation; it does not establish that the published scientific claim is false.", "no_adaptive_changes": "Do not widen tolerance, change seeds, input precision, metrics, data size, initialization or refinement in response to results. Preserve any failed or superseded run." }, "preflight": [ "Exact upstream source and lock hashes", "Exact pinned core runtime and AVX2 activation", "All13 publisher input hashes and expected tensors", "All five complete original configurations and disjoint halves", "Upstream test_method.py and test_evaluation.py pass", "Fresh output directory and no stale matching results" ], "completion_requirements": [ "Every predetermined result file exists with complete finite metrics and correct configuration", "Five original seeds retained and aggregated", "Full input/source hashes unchanged after execution", "Independent comparison to frozen public references", "Publish numerical results, evidence and clear limits on GitHub and Cloudflare" ], "scope_limits": [ "Only one published configuration is reproduced, not the whole paper.", "Released embeddings replay alignment; they do not independently authenticate pretraining or re-extract image/text features.", "The five-photo direct-matching experiment remains a separate exploratory result.", "Model-choice controls, separate source-overlap controls and a prospective fresh collection remain later phases and are not completed by this baseline.", "New collection requires actual later human photographs and descriptions; none may be fabricated or relabeled as a fresh capture." ], "bootstrap_receipt_sha256": "8ae61f5d484ac19b505f037bab2ad3fc076d2b2429bc4deb1a10167562a1789f", "expected_input_count": 13, "expected_source_file_count": 228, "expected_training_counts": { "x": 41391, "y": 41392 }, "expected_validation_count": 40504 }