Recompute the OWT unigram surprisal of Provo

README.md

@@ -228,7 +228,7 @@ Start from `tags.parquet`: one row per (`tag`, `predictor`, file), with the
 model, method, configuration, storage class, layer and provenance of every
 cell decoded, so a cell can be found by model, method and configuration.
 
-- 2,596 cells across 58 corpora, 280,034,199 scalar rows
+- 2,596 cells across 58 corpora, 280,036,007 scalar rows
 - 25 methods, 26 models and frequency tables
 
 ## Layout
@@ -283,7 +283,7 @@ A `positional` estimator conditions on a prefix ending at a token position and r
 | `unit_rule` | cells |
 |---|---|
 | `spacy_sentence` | 26 |
-| `whitespace_continuation` | 403 |
+| `whitespace_continuation` | 406 |
 
 ## Word values derived from token tables
 
@@ -351,6 +351,8 @@ Each recompute wave was checked table by table against the corpus loader's unit
 | file | verdict | cells |
 |---|---|---|
 | `verification/meco-nlp-row.json` | pass | 4 |
+| `verification/owt-unigram.json` | check | 2 |
+| `verification/owt-unigram.json` | pass | 1 |
 | `verification/wave1-tokens.json` | check | 406 |
 | `verification/wave1-tokens.json` | pass | 4 |
 | `verification/wave2.json` | check | 272 |
@@ -360,7 +362,7 @@ Each recompute wave was checked table by table against the corpus loader's unit
 |---|---|
 | manifest identity lacks storage_class | 405 |
 | manifest layer is None | 405 |
-| words empty or containing whitespace | 277 |
+| words empty or containing whitespace | 279 |
 | context scope story (layer 1 conditions on the whole stimulus) | 49 |
 | BOS disabled | 42 |
 | context scope story | 28 |
@@ -387,7 +389,7 @@ Each recompute wave was checked table by table against the corpus loader's unit
 |---|---|
 | absent | 1 |
 | carried | 35 |
-| exported | 1319 |
+| exported | 1322 |
 | quarantined | 5 |
 
 | `status` | `reason_class` | tables |
@@ -400,7 +402,7 @@ Each recompute wave was checked table by table against the corpus loader's unit
 | carried | missing | 2 |
 | exported | complete | 863 |
 | exported | expected_refusal | 2 |
-| exported | not planned | 454 |
+| exported | not planned | 457 |
 | quarantined | held_decision | 5 |
 
 | quarantined `method` | tables |
@@ -411,27 +413,28 @@ Each recompute wave was checked table by table against the corpus loader's unit
 
 ## What changed in this release
 
-Built on the published store at commit `1227e0e3e66b85588bf1ceee4fc3f524b6901a1f`; every published cell this release does not replace is carried over unchanged.
+Built on the published store at commit `3ecb7c92de0e312b0d543259ce591e6552eabf87`; every published cell this release does not replace is carried over unchanged.
 
 | wave | cells exported |
 |---|---|
 | meco-nlp-row | 4 |
+| owt-unigram | 3 |
 | wave1-tokens | 410 |
 | wave2 | 903 |
 
-- 1317 published cells replaced by a recomputed table of the same tag
+- 1320 published cells replaced by a recomputed table of the same tag
 - 0 published files deleted because their cell was recomputed without the array they held
-- 1910 files written
+- 1914 files written
 - new: `tokens/` (token-grain tables), `characters/` (character-grain tables), `coverage.parquet`, `verification/`, and the storage columns of the word tables and the index
 - array keys without a store folder, left in the pickles only: `orth_similarity_samples`, `pos_match_samples`, `sampled_words`, `similarity_samples_tokens`
 
 ### Carried cells on an older unit grid
 
-128 cells this release carries were computed against an earlier version of their corpus, so their `stimulus_id` and `unit_index` do not line up with the cells this release stages into the same table. Join them to a recomputed predictor only after checking the units match. `tags.parquet` gives each cell's `n_stimuli` and `n_rows`, and a recomputed cell of the same corpus gives the current shape.
+126 cells this release carries were computed against an earlier version of their corpus, so their `stimulus_id` and `unit_index` do not line up with the cells this release stages into the same table. Join them to a recomputed predictor only after checking the units match. `tags.parquet` gives each cell's `n_stimuli` and `n_rows`, and a recomputed cell of the same corpus gives the current shape.
 
 | table | cells on an older grid |
 |---|---|
-| `dundee.parquet` | 19 |
+| `dundee.parquet` | 18 |
 | `meco_de.parquet` | 2 |
 | `meco_ee.parquet` | 6 |
 | `meco_fi.parquet` | 5 |
@@ -442,7 +445,7 @@ Built on the published store at commit `1227e0e3e66b85588bf1ceee4fc3f524b6901a1f
 | `meco_l2_w2.parquet` | 10 |
 | `meco_tr.parquet` | 5 |
 | `onestop.parquet` | 16 |
-| `provo.parquet` | 15 |
+| `provo.parquet` | 14 |
 | `sbsat.parquet` | 20 |
 | `szewczyk_n400.parquet` | 8 |
 

coverage.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c5298aaaafd15159121bff709cb271bcecc15437365e62b33b799297b33d0bf9
-size 16770
+oid sha256:97499a016f508705b7b13a5e9e8325d48d548fad37ef56b18f35b02273d16657
+size 16815

dundee.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f10ff9131c9d976b461907c515667e0cd08a478139f72c526499dd55a4c0faa
-size 40845662
+oid sha256:24268807b37c745eed3378e3dc9f432e72dbff1dc779098329d72caf20e825f8
+size 40915707

geco.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d2001b86daa646dac72dfad567aa9a0857d0628013f108ae5924bb2dbbfabd54
-size 46633317
+oid sha256:76a34b6c35b3a05b8636699cc4599eedfa327afc62da735565efa13864be6cb7
+size 46623563

history.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d95135418f256a0ddd34b1547359b970aa9ded204ef9a52afbffb2cf9702bb9a
-size 29988
+oid sha256:435fd5ca82efd87568924c386850ec1ecd2660a805d0dc97b5c851e78eed095f
+size 30037

manifests/owt_unigram_surprisal_dundee.json

@@ -0,0 +1,60 @@
+{
+  "identity": {
+    "schema_version": 1,
+    "method": "unigram_surprisal",
+    "storage_class": "unit_bound",
+    "layer": 1,
+    "options": {
+      "model_id": "owt",
+      "beam_width": 5,
+      "n_samples": 64,
+      "max_tokens": 10,
+      "alpha": 1.0,
+      "seed": null,
+      "add_bos": true,
+      "dataset_language": "en",
+      "language_override": null,
+      "chunk_size": null,
+      "word_starters": [],
+      "top_p": 0.9,
+      "context_scope": "stimulus",
+      "unit": "words",
+      "window_policy": "reference",
+      "segmentation": "whitespace",
+      "model_revision": null,
+      "store_distribution": false,
+      "model_identity": null
+    },
+    "model": {
+      "source": "owt",
+      "table": "unigram_counts.tsv",
+      "sha256": "c68e8ac6d3511a7dcc6438ba74935eb0465118a3c3f045313d17ef6087a400d3"
+    },
+    "corpus_sha256": "7e634bbd21030e4d5386193726ebf74957ae28f4555911a655154211cd9f8d88",
+    "n_stimuli": 20,
+    "source_sha256": "c149055d0edd68870401df6c1de2aaae75378e8b47814f51a800526489c773da",
+    "versions": {
+      "python": "3.12.13",
+      "replica": "0.0.1",
+      "numpy": "2.4.4",
+      "torch": "2.11.0",
+      "transformers": "4.57.6",
+      "tokenizers": "0.22.2",
+      "wordfreq": "3.1.1",
+      "wordsprobability": "0.17",
+      "genlm-bytes": "0.2.0"
+    },
+    "unit_rule": "whitespace_continuation",
+    "unit_layer": "words"
+  },
+  "status": "completed",
+  "payload_sha256": "2935e1f88b0240fd6b508839afa05b001863098d3ec96a948ce4aee99e768bcb",
+  "environment": {
+    "code_commit": "51410e655a99691d90fcf3bc48196ad711f1c9b2",
+    "code_dirty": false,
+    "device": "cpu",
+    "model_device": null,
+    "cuda": null,
+    "computed_at": "2026-09-30T10:26:27+00:00"
+  }
+}

manifests/owt_unigram_surprisal_geco.json

@@ -0,0 +1,60 @@
+{
+  "identity": {
+    "schema_version": 1,
+    "method": "unigram_surprisal",
+    "storage_class": "unit_bound",
+    "layer": 1,
+    "options": {
+      "model_id": "owt",
+      "beam_width": 5,
+      "n_samples": 64,
+      "max_tokens": 10,
+      "alpha": 1.0,
+      "seed": null,
+      "add_bos": true,
+      "dataset_language": "en",
+      "language_override": null,
+      "chunk_size": null,
+      "word_starters": [],
+      "top_p": 0.9,
+      "context_scope": "stimulus",
+      "unit": "words",
+      "window_policy": "reference",
+      "segmentation": "whitespace",
+      "model_revision": null,
+      "store_distribution": false,
+      "model_identity": null
+    },
+    "model": {
+      "source": "owt",
+      "table": "unigram_counts.tsv",
+      "sha256": "c68e8ac6d3511a7dcc6438ba74935eb0465118a3c3f045313d17ef6087a400d3"
+    },
+    "corpus_sha256": "29d759acef49c2eb2ab44d396f83853833e24ba5fcbaef7c50132819fc6aab95",
+    "n_stimuli": 588,
+    "source_sha256": "c149055d0edd68870401df6c1de2aaae75378e8b47814f51a800526489c773da",
+    "versions": {
+      "python": "3.12.13",
+      "replica": "0.0.1",
+      "numpy": "2.4.4",
+      "torch": "2.11.0",
+      "transformers": "4.57.6",
+      "tokenizers": "0.22.2",
+      "wordfreq": "3.1.1",
+      "wordsprobability": "0.17",
+      "genlm-bytes": "0.2.0"
+    },
+    "unit_rule": "whitespace_continuation",
+    "unit_layer": "words"
+  },
+  "status": "completed",
+  "payload_sha256": "a3ca8141a7ac49227ae1be39e1c9afdb4de093be32456b392ed8b7bfbd2aa22d",
+  "environment": {
+    "code_commit": "51410e655a99691d90fcf3bc48196ad711f1c9b2",
+    "code_dirty": false,
+    "device": "cpu",
+    "model_device": null,
+    "cuda": null,
+    "computed_at": "2026-09-30T10:30:16+00:00"
+  }
+}

manifests/owt_unigram_surprisal_provo.json

@@ -0,0 +1,60 @@
+{
+  "identity": {
+    "schema_version": 1,
+    "method": "unigram_surprisal",
+    "storage_class": "unit_bound",
+    "layer": 1,
+    "options": {
+      "model_id": "owt",
+      "beam_width": 5,
+      "n_samples": 64,
+      "max_tokens": 10,
+      "alpha": 1.0,
+      "seed": null,
+      "add_bos": true,
+      "dataset_language": "en",
+      "language_override": null,
+      "chunk_size": null,
+      "word_starters": [],
+      "top_p": 0.9,
+      "context_scope": "stimulus",
+      "unit": "words",
+      "window_policy": "reference",
+      "segmentation": "whitespace",
+      "model_revision": null,
+      "store_distribution": false,
+      "model_identity": null
+    },
+    "model": {
+      "source": "owt",
+      "table": "unigram_counts.tsv",
+      "sha256": "c68e8ac6d3511a7dcc6438ba74935eb0465118a3c3f045313d17ef6087a400d3"
+    },
+    "corpus_sha256": "2e84d19d8d281c7337c5d39839bad1072854fd00cd94caf737dbbcc01caa398e",
+    "n_stimuli": 55,
+    "source_sha256": "c149055d0edd68870401df6c1de2aaae75378e8b47814f51a800526489c773da",
+    "versions": {
+      "python": "3.12.13",
+      "replica": "0.0.1",
+      "numpy": "2.4.4",
+      "torch": "2.11.0",
+      "transformers": "4.57.6",
+      "tokenizers": "0.22.2",
+      "wordfreq": "3.1.1",
+      "wordsprobability": "0.17",
+      "genlm-bytes": "0.2.0"
+    },
+    "unit_rule": "whitespace_continuation",
+    "unit_layer": "words"
+  },
+  "status": "completed",
+  "payload_sha256": "97b0803e48ddf4db0cc147ab36d42276967d7a186f0636f1b8d0ef8faa5c3fbc",
+  "environment": {
+    "code_commit": "51410e655a99691d90fcf3bc48196ad711f1c9b2",
+    "code_dirty": false,
+    "device": "cpu",
+    "model_device": null,
+    "cuda": null,
+    "computed_at": "2026-09-30T09:53:30+00:00"
+  }
+}

provo.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bc2059c409e9f11bd68ae88158e459988aef91d462323db3eadf13c854127bfd
-size 2735079
+oid sha256:6c32b98960efe3d541dd67c619d296d112df6baea9079d3da7b6e8efd55cd96e
+size 2744514

tags.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f368133c2a023da55773dce5ac9b4e7234f4f5b3ce18996a1c4dbb3d6e7787c8
-size 73859
+oid sha256:33815219bf39106ccc7d36b4198e7d12071308c5e7f17b11bc6b42651f497b91
+size 73973