Upload folder using huggingface_hub

README.md

@@ -0,0 +1,3 @@
+# Exact experiment data archive
+
+Archive of exact NumPy caches, split metadata and available generator/source files. No re-sampling or train/test relabeling is performed. Split metadata records seeds, tokenizer/source revisions and packing. A training-data evaluation remains a training-data evaluation; independently sampled synthetic tests are distinct from held-out natural-language documents. dataset_manifest.json records checksums; experiment_configs/ records the consuming experiments.

construction.json

@@ -0,0 +1,9 @@
+{
+  "source_report_sha256": "d7f3a988eb899cb13ed596fcb74c397c594d85257250d40c85b3989ac298514a",
+  "tokens": 199999488,
+  "seed": 40,
+  "source_context_length": 1024,
+  "training_context_length": 256,
+  "generation_batch_size": 512,
+  "packing_version": "empirical-trigram-subwindows-v1"
+}

dataset_audit.json

@@ -0,0 +1,23 @@
+{
+  "tokens": 199999488,
+  "sequences": 781248,
+  "shape": [
+    781248,
+    257
+  ],
+  "train_sha256": "0fbe4a5af5102983aa7c5e1ec276db7fe34992c5f8c646a3e21cc81e088d8a09",
+  "train_metadata_sha256": "58fd29f4e293327604ecfb7d2093041a54ad6d31cd97353727a12ab5d6539a0d",
+  "source_report_sha256": "d7f3a988eb899cb13ed596fcb74c397c594d85257250d40c85b3989ac298514a",
+  "packing_verified_all_windows": true,
+  "source_sampling_audit": {
+    "trigram_predictions": 199804176,
+    "source_cross_entropy_nats": 3.5715664398945886,
+    "expected_trigram_entropy_nats": 3.5714021743012294,
+    "unsupported_triples": 0,
+    "support_check": "Each draw is indexed into an empirical-support row; no fallback",
+    "sample_sha256": "ef2c0a9645bcf5f758a97274b5b62473700a1f71b545a882fc7036bf1f831565",
+    "elapsed_seconds": 272.5424264149624
+  },
+  "evaluation_kind": "training_sample",
+  "note": "Exact training cache and source provenance; full fitted sparse probability tables remain in the original source directory."
+}

experiment_configs/trigram_components_001_2b_schedule_200m_dmodel256.json

@@ -0,0 +1,60 @@
+{
+  "data_dir": "outputs/research/natural_language_scaling/trigram_components_001_200m_data",
+  "output_dir": "outputs/research/natural_language_scaling/trigram_components_001_2b_schedule_200m_dmodel256",
+  "tokenizer": "eoinf/pile_tokenizer_4096",
+  "tokenizer_revision": "155b268a49b69bac2543a3eb0acb635cef3dc4b0",
+  "train_tokens": 199999488,
+  "evaluate_on_training_data": true,
+  "vocab_size": 4096,
+  "context_length": 256,
+  "widths": [
+    256
+  ],
+  "num_layers": 2,
+  "num_heads": 1,
+  "seeds": 1,
+  "training_tokens": 1999998976,
+  "micro_batch_size": 16,
+  "gradient_accumulation": 1,
+  "eval_batch_size": 1,
+  "evaluation_windows": 128,
+  "checkpoint_count": 20,
+  "learning_rate": 0.001,
+  "warmup_steps": 100,
+  "final_lr_ratio": 0.01,
+  "weight_decay": 0.0,
+  "grad_clip": 1.0,
+  "torch_threads": 1,
+  "device": "cuda",
+  "workers": 1,
+  "mlp_width_multiplier": 1,
+  "training_log_interval": 50,
+  "training_ewma_beta": 0.99,
+  "data_kind": "synthetic_trigram",
+  "plot_title": "Entropy-matched trigrams: components 001",
+  "stop_after_tokens": 199999488,
+  "source_generator": "power_decay_empirical_support_trigram",
+  "source_dir": "outputs/research/natural_language_scaling/trigram_components_001_source",
+  "dataset_public": true,
+  "source_rank": 1,
+  "source_alpha": 1.0,
+  "source_seed": 20261103,
+  "target_entropy_nats": 3.5714023607206284,
+  "reference_dir": "outputs/research/natural_language_scaling/trigram_reference_inputs",
+  "reference_git_branch": "paperspace-data-trigram-reference-v1",
+  "reference_git_revision": "3793aa8283d295997ef470fe288cc7806bbd9297",
+  "reference_manifest_sha256": "ad7a97d6ba5bac15151bd954513cc7f1d2941482ff9da13ca43b6f6ec1181d91",
+  "reference_complete_sha256": "535c57b1e3ef35730e26c68987d44cf48df8a5f383387061337dbc7f841ef765",
+  "preparation_code_sha256": {
+    "scripts/research/prepare_power_decay_trigram.py": "7ac59348d9c9765f62fab2d700fa88cdf1bab515795a6a8bde39b65ccb505dc1",
+    "scripts/research/sample_power_decay_trigram.py": "10cf767b6f07b9cd11589ac1a6f87284574ebed7280da92ecdb526db27cf7943",
+    "scripts/research/prepare_power_decay_trigram_training.py": "8470a67b6ec6424b231255fb1ab299294a77d8f9921f0f44ddcdc856dd76e842",
+    "src/synthetic_trigrams/empirical_trigram/power_decay.py": "857df64c53364c3deb9ae4efe15db3fe7176d0bf841cb6f508a327c8f0b12fa9",
+    "src/synthetic_trigrams/empirical_trigram/feature_variants.py": "0427f2ec5aae333fc14d04e910e4b7dda8c38ca8f05fda85f7c888e6ca15c14c",
+    "src/synthetic_trigrams/natural_language_scaling/prepare_hf_training_grid.py": "8f18165074dcffa1d7bdea2359f22a83e918597d8a9fe7a3e2479f19c008f7ea"
+  },
+  "source_report_sha256": "d7f3a988eb899cb13ed596fcb74c397c594d85257250d40c85b3989ac298514a",
+  "expected_data_metadata_sha256": {
+    "train": "58fd29f4e293327604ecfb7d2093041a54ad6d31cd97353727a12ab5d6539a0d"
+  }
+}

source_report.json

@@ -0,0 +1,83 @@
+{
+  "beta": 5.521713056250135,
+  "blocks": 4096,
+  "occurrence_mass": 1.0,
+  "conditional_entropy_nats": 3.5714021743012294,
+  "pair_marginal_l1": 4.525853572072085e-09,
+  "reference": {
+    "vocab": 4096,
+    "original_source": {
+      "identity": {
+        "fit_report_sha256": "7dafa8fe382295e77ce4ed32683ebe56ce83f0b9d25beed2a4410ee7d73c258e",
+        "token_sha256": "f4ec26137d7487429f68a6e977d8386479f205001a6ec7c424bb90f21e6869c9",
+        "suffix_sha256": "b3487d310803cd504260137665a1db1361c0d5c0b24e747747a2673dc2e266f2",
+        "reconstruction": "merged fourgram prefix counts plus final triples; circular reverse histories v2"
+      },
+      "entropy": 3.5714023585974424,
+      "reference_entropy": 3.571402358583158,
+      "observed_pairs": 5910900,
+      "occurrences": 1999999998,
+      "history_boundary_triples": 2,
+      "source_token_count": 2000000000,
+      "entropy_difference_from_archive": 1.4284573524037114e-11,
+      "note": "Target counts reconstructed from the available 2B natural stream prefix. Matching supported-pair count and near-matching entropy are consistency checks, not proof of byte-identical original source counts. Reverse histories add two wraparound triples to make empirical pair weights exactly stationary."
+    },
+    "reference_sha256": "535c57b1e3ef35730e26c68987d44cf48df8a5f383387061337dbc7f841ef765",
+    "observed_trigrams": 132810622,
+    "observed_contexts": 5910900,
+    "occurrences": 2000000002,
+    "added_supported_triples": [
+      [
+        480,
+        2879,
+        294
+      ],
+      [
+        2879,
+        294,
+        2
+      ],
+      [
+        294,
+        2,
+        1
+      ],
+      [
+        2,
+        1,
+        1465
+      ]
+    ],
+    "support_matches_original_empirical": true,
+    "original_occurrences": 1999999998,
+    "pair_frequency_tvd_upper_bound": 1.999999998e-09,
+    "conditional_entropy_nats": 3.5714023607206284,
+    "bigram_conditional_entropy_nats": 4.87638457764605,
+    "additional_information_nats": 1.3049822169254215,
+    "deterministic_contexts": 1457930,
+    "deterministic_occurrence_mass": 0.0028618434971381564,
+    "pair_flow_count_error": 0,
+    "note": "Removed two artificial wrap edges, then added one count along an existing-support return path. Exact stationary reference; probabilities differ minutely from the finite-stream empirical source."
+  },
+  "identity": {
+    "reference": "/notebooks/synthetic_trigrams/outputs/research/natural_language_scaling/trigram_reference_inputs/original",
+    "reference_sha256": "535c57b1e3ef35730e26c68987d44cf48df8a5f383387061337dbc7f841ef765",
+    "rank": 1,
+    "alpha": 1.0,
+    "seed": 20261103,
+    "tolerance": 1e-08,
+    "implementation": "sparse-middle-blocks-v2",
+    "feature_bank_rank": 128,
+    "context_features": "independent",
+    "context_frequency_exponent": 0.0,
+    "frequency_pivot_count": 18063.0,
+    "frequency_multiplier_bounds": [
+      0.5,
+      2.0
+    ]
+  },
+  "target_entropy_nats": 3.5714023607206284,
+  "additional_information_nats": 1.3049824033448205,
+  "probabilities_directory": "outputs/research/natural_language_scaling/trigram_components_001_source/fits/beta_5.5217130563",
+  "note": "Exact empirical support; stationary pair marginals and total conditional information matched. Per-context information is not fitted. Coefficients follow power decay before masking and balancing; no exact final spectrum claim."
+}

train.json

@@ -0,0 +1,22 @@
+{
+  "source": "synthetic_empirical_support_power_decay_trigram",
+  "source_revision": "d7f3a988eb899cb13ed596fcb74c397c594d85257250d40c85b3989ac298514a",
+  "source_report_sha256": "d7f3a988eb899cb13ed596fcb74c397c594d85257250d40c85b3989ac298514a",
+  "source_rows_sha256": "ef2c0a9645bcf5f758a97274b5b62473700a1f71b545a882fc7036bf1f831565",
+  "tokenizer_name": "eoinf/pile_tokenizer_4096",
+  "tokenizer_revision": "155b268a49b69bac2543a3eb0acb635cef3dc4b0",
+  "eos_token_id": 2,
+  "seed": 40,
+  "generation_batch_size": 512,
+  "source_conditional_entropy_nats": 3.5714021743012294,
+  "training_window_entropy_nats": 3.576499761814295,
+  "evaluation_kind": "training_sample",
+  "note": "Same 1025-to-257 source-row packing as historical empirical trigrams; synthetic source changes. Training-sample evaluation is not held out.",
+  "tokens": 199999488,
+  "cache_entries": 781248,
+  "tokenizer_size": 4096,
+  "packing": "Subwindows of independent source rows; 257 tokens and 256 scored targets per window",
+  "row_length": 1025,
+  "windows_per_source_row": 4,
+  "context_length": 256
+}

train.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fbe4a5af5102983aa7c5e1ec276db7fe34992c5f8c646a3e21cc81e088d8a09
+size 401561600