AvrovaDonz commited on
Commit
ffb7e64
·
verified ·
1 Parent(s): cd3a000

docs(dataset): record the measured record mix and the no-verification caveat

Browse files
components/google-math/README.md CHANGED
@@ -22,6 +22,11 @@ school-level difficulty. Those pairs are **not research mathematics**, and the s
22
  projection of the paper's curriculum rather than the paper's dataset. Every record carries
23
  `content_origin` so consumers can separate verbatim source text from `generator_output`.
24
 
 
 
 
 
 
25
  ## Included sources
26
 
27
  Each source is pinned to a commit or dataset revision, and each ships math text under Apache-2.0
@@ -90,6 +95,7 @@ The short version:
90
  measured yield.
91
  - `audit/excluded-duplicates.jsonl`, `audit/packing.json`, `audit/source-scan.json`.
92
  - `audit/validation-evidence.json`: the independent decode and round-trip evidence described below.
 
93
  - `statistics.json`: measured records, groups, splits, characters and packed counts, per source.
94
  - `manifest.json`: SHA256 and byte size for every delivered file except itself and validation.json.
95
  - `packed/`: mmla-type2-packed-v1 arrays, manifest, row attribution and per-record token counts.
@@ -163,10 +169,11 @@ boundary masks, and re-encodes every record's token stream against `data/*.jsonl
163
  `all_packed_tokens_reencoded_source_shift_padding_masks_and_attribution`, `statistics_match_artifacts`).
164
 
165
  `tools/verify_google_math.py` is the independent check against the delivered files, and its output is
166
- [`audit/validation-evidence.json`](audit/validation-evidence.json). It decoded 96 packed rows spread over
167
- every shard of all three splits, and for each row (a) confirmed every row slice equals the re-encoding of
168
- the corresponding span of `[BOS] + SentencePiece(text) + [EOS]` for its attributed source record and
169
- (b) confirmed the decoded text equals the decoded source text modulo whitespace. It recorded 8 literal
 
170
  unified diffs of a decoded row against its source record; every diff had 0 lines, whitespace identical.
171
  It round-tripped 3 Parquet rows (indices 0, 239,658 and 479,315) back to their JSONL records with exact
172
  match, and confirmed a `formal_lean_proof`, a `formal_lean_statement`, a `generated_qa_pair`, a
@@ -174,20 +181,34 @@ match, and confirmed a `formal_lean_proof`, a `formal_lean_statement`, a `genera
174
  matching `text_sha256`. Split separation was recomputed from the data rather than trusted: 3,403 groups
175
  in the audit and 3,403 in the data, no group spanning two splits, train/dev/test at 3,333/35/35 groups.
176
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
177
  ## Publication
178
 
179
- The component is published as Hugging Face dataset pull request
180
- [MMLA-ORG/Cayley-Math discussion #5](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/discussions/5),
181
- on ref `refs/pr/5`, head commit `1894ffe2dc7f7f59abfb830f91e79738b1eb11d3`. It carries the whole
182
- component plus the top-level `README.md` config and Collections update, and it has not been merged into
183
- `main`.
184
 
185
  A token with `write` scope or org-admin rights is required to merge it. The token used for this build can
186
  read the repository and open pull requests, but a direct commit to `main` is rejected with
187
- `403 Forbidden: You have read access but not the required permissions for this operation`. Merging that
188
  pull request, or re-running `hf upload` without `--create-pr` once such a token is available, completes
189
- the submission. Earlier pull requests #3 and #4 are superseded by #5 and were left open rather than
190
- closed, because closing them was outside this task's scope.
191
 
192
  ## Limitations
193
 
 
22
  projection of the paper's curriculum rather than the paper's dataset. Every record carries
23
  `content_origin` so consumers can separate verbatim source text from `generator_output`.
24
 
25
+ **This corpus is mostly seeded generator output, not research mathematics.** `generated_qa_pair` records
26
+ dominate the record count, the formal subset is a small minority, and **no Lean compilation and no proof
27
+ verification was performed on any subset**, so no record here has a checked proof or a checked statement.
28
+ `statistics.json` carries the measured counts, and the component README states them in full.
29
+
30
  ## Included sources
31
 
32
  Each source is pinned to a commit or dataset revision, and each ships math text under Apache-2.0
 
95
  measured yield.
96
  - `audit/excluded-duplicates.jsonl`, `audit/packing.json`, `audit/source-scan.json`.
97
  - `audit/validation-evidence.json`: the independent decode and round-trip evidence described below.
98
+ - `audit/publish-check.json`: the second publish gate, measured from the delivered files after the build.
99
  - `statistics.json`: measured records, groups, splits, characters and packed counts, per source.
100
  - `manifest.json`: SHA256 and byte size for every delivered file except itself and validation.json.
101
  - `packed/`: mmla-type2-packed-v1 arrays, manifest, row attribution and per-record token counts.
 
169
  `all_packed_tokens_reencoded_source_shift_padding_masks_and_attribution`, `statistics_match_artifacts`).
170
 
171
  `tools/verify_google_math.py` is the independent check against the delivered files, and its output is
172
+ [`audit/validation-evidence.json`](audit/validation-evidence.json). It decoded 96 packed rows, covering
173
+ 68 of the 70 train shards plus the single dev and test shards, and for each row (a) confirmed every row
174
+ slice equals the re-encoding of the corresponding span of `[BOS] + SentencePiece(text) + [EOS]` for its
175
+ attributed source record and (b) confirmed the decoded text equals the decoded source text modulo
176
+ whitespace. It recorded 8 literal
177
  unified diffs of a decoded row against its source record; every diff had 0 lines, whitespace identical.
178
  It round-tripped 3 Parquet rows (indices 0, 239,658 and 479,315) back to their JSONL records with exact
179
  match, and confirmed a `formal_lean_proof`, a `formal_lean_statement`, a `generated_qa_pair`, a
 
181
  matching `text_sha256`. Split separation was recomputed from the data rather than trusted: 3,403 groups
182
  in the audit and 3,403 in the data, no group spanning two splits, train/dev/test at 3,333/35/35 groups.
183
 
184
+ [`audit/publish-check.json`](audit/publish-check.json) is a second publish gate with its own code rather
185
+ than this component's tools. It agrees with every total above (480,584 records, 3,403 groups, 78,364,888
186
+ characters, and the packed counts for all three splits), decodes 12 packed rows across all three splits
187
+ and compares 263 attributed segments. All 263 row slices matched their canonical token slices by exact
188
+ identity, and all 248 segments that cover a whole record decoded to text identical to the source record
189
+ under whitespace normalisation and re-encoded to the identical token sequence. The other 15 segments are
190
+ row-boundary fragments that begin or end inside a SentencePiece piece, so they cannot re-encode
191
+ fragment-by-fragment; each one's decoded text was still located verbatim inside its attributed record.
192
+ It matched 5 of 5 Parquet round-trips exactly and recomputed group separation from `data/*.jsonl.gz`
193
+ directly, finding 0 groups spanning two splits.
194
+
195
+ The record mix is deliberately stated because it is skewed: `generated_qa_pair` records are 99.31% of
196
+ records but 45.57% of characters, while formal Lean records are 0.63% of records and 54.22% of characters,
197
+ and all non-generator records together are 0.69% of records. A consumer wanting formal or prose
198
+ mathematics should filter on `record_type` rather than read the corpus whole.
199
+
200
  ## Publication
201
 
202
+ The component is delivered as Hugging Face dataset pull request
203
+ [MMLA-ORG/Cayley-Math discussion #3](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/discussions/3)
204
+ on ref `refs/pr/3`. It carries the whole component plus the top-level `README.md` config and Collections
205
+ update, and it has not been merged into `main`. Consult that discussion for the current merge state.
 
206
 
207
  A token with `write` scope or org-admin rights is required to merge it. The token used for this build can
208
  read the repository and open pull requests, but a direct commit to `main` is rejected with
209
+ `403 Forbidden: You have read access but not the required permissions for this operation`. Merging the
210
  pull request, or re-running `hf upload` without `--create-pr` once such a token is available, completes
211
+ the submission.
 
212
 
213
  ## Limitations
214
 
components/google-math/manifest.json CHANGED
@@ -13,8 +13,8 @@
13
  },
14
  {
15
  "path": "README.md",
16
- "bytes": 18839,
17
- "sha256": "f099faeb03876d81ec079d556a51b16cb93a0ae9185842c19bc031b439f1dd91"
18
  },
19
  {
20
  "path": "audit/excluded-duplicates.jsonl",
@@ -1263,8 +1263,8 @@
1263
  },
1264
  {
1265
  "path": "tools/build_google_math.py",
1266
- "bytes": 60682,
1267
- "sha256": "b4fd0ed309814edee92d83e4a20f2bae1ff5365600b291462ffb22a59e41f761"
1268
  },
1269
  {
1270
  "path": "tools/generate_mathematics_dataset_sample.py",
 
13
  },
14
  {
15
  "path": "README.md",
16
+ "bytes": 20593,
17
+ "sha256": "b3c98cc1005b5263dbe2bad37aafba811bf69ac6f268a918bcc851fe6f83af6d"
18
  },
19
  {
20
  "path": "audit/excluded-duplicates.jsonl",
 
1263
  },
1264
  {
1265
  "path": "tools/build_google_math.py",
1266
+ "bytes": 61871,
1267
+ "sha256": "8e72001f8b24f5d188c73da7f4e9e29057300e5fa279e8d370f297375f6202d4"
1268
  },
1269
  {
1270
  "path": "tools/generate_mathematics_dataset_sample.py",
components/google-math/tools/build_google_math.py CHANGED
@@ -930,6 +930,11 @@ school-level difficulty. Those pairs are **not research mathematics**, and the s
930
  projection of the paper's curriculum rather than the paper's dataset. Every record carries
931
  `content_origin` so consumers can separate verbatim source text from `generator_output`.
932
 
 
 
 
 
 
933
  ## Included sources
934
 
935
  Each source is pinned to a commit or dataset revision, and each ships math text under Apache-2.0
@@ -1075,19 +1080,53 @@ match, and confirmed a `formal_lean_proof`, a `formal_lean_statement`, a `genera
1075
  matching `text_sha256`. Split separation was recomputed from the data rather than trusted: 3,403 groups
1076
  in the audit and 3,403 in the data, no group spanning two splits, train/dev/test at 3,333/35/35 groups.
1077
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1078
  ## Limitations
1079
 
1080
  - No Lean compilation, proof checking, Comparator run or mathematical verification of any kind.
 
 
 
1081
  - PDF prose proofs are noisy pypdf text; symbol and ligature fidelity is not guaranteed.
1082
  - The generator subset is a seeded sample of one curriculum, not the paper's full dataset, and its
1083
- questions are school-level rather than research mathematics.
 
 
1084
  - `formal-conjectures` upstream states that some content derives from third-party sources (Wikipedia,
1085
  MathOverflow and OEIS under CC-BY-SA 4.0). Per-file reference URLs are retained inside the Lean text; no
1086
  separate per-file license audit was performed.
1087
- - Near-duplicate detection, Phase 1 overlap analysis and editorial review were not performed.
 
1088
  - The SentencePiece model normalises runs of spaces, so packed text decodes whitespace-normalised rather
1089
  than byte-identically to the sources.
1090
- - Only problems that upstream published are present, so several subsets are selection-biased.
 
 
1091
 
1092
  ## Statistics
1093
 
 
930
  projection of the paper's curriculum rather than the paper's dataset. Every record carries
931
  `content_origin` so consumers can separate verbatim source text from `generator_output`.
932
 
933
+ **This corpus is mostly seeded generator output, not research mathematics.** `generated_qa_pair` records
934
+ dominate the record count, the formal subset is a small minority, and **no Lean compilation and no proof
935
+ verification was performed on any subset**, so no record here has a checked proof or a checked statement.
936
+ `statistics.json` carries the measured counts, and the component README states them in full.
937
+
938
  ## Included sources
939
 
940
  Each source is pinned to a commit or dataset revision, and each ships math text under Apache-2.0
 
1080
  matching `text_sha256`. Split separation was recomputed from the data rather than trusted: 3,403 groups
1081
  in the audit and 3,403 in the data, no group spanning two splits, train/dev/test at 3,333/35/35 groups.
1082
 
1083
+ [`audit/publish-check.json`](audit/publish-check.json) is a third check written by an independent publish
1084
+ gate with its own code rather than this component's tools. It agrees with every total above (480,584
1085
+ records, 3,403 groups, 78,364,888 characters), decodes 12 packed rows covering 248 whole-record segments
1086
+ with 248/248 token-identical and whitespace-identical, matches 5/5 Parquet round-trips, and reports zero
1087
+ groups spanning two splits.
1088
+
1089
+ The record mix is deliberately stated because it is skewed: `generated_qa_pair` records are 99.31% of
1090
+ records but 45.57% of characters, while formal Lean and prose records are 0.69% of records and 54.43% of
1091
+ characters. A consumer wanting formal or prose mathematics should filter on `record_type` rather than
1092
+ read the corpus whole.
1093
+
1094
+ ## Publication
1095
+
1096
+ The component is published as Hugging Face dataset pull request
1097
+ [MMLA-ORG/Cayley-Math discussion #5](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/discussions/5),
1098
+ on ref `refs/pr/5`, head commit `1894ffe2dc7f7f59abfb830f91e79738b1eb11d3`. It carries the whole
1099
+ component plus the top-level `README.md` config and Collections update, and it has not been merged into
1100
+ `main`.
1101
+
1102
+ A token with `write` scope or org-admin rights is required to merge it. The token used for this build can
1103
+ read the repository and open pull requests, but a direct commit to `main` is rejected with
1104
+ `403 Forbidden: You have read access but not the required permissions for this operation`. Merging that
1105
+ pull request, or re-running `hf upload` without `--create-pr` once such a token is available, completes
1106
+ the submission. Earlier pull requests #3 and #4 are superseded by #5 and were left open rather than
1107
+ closed, because closing them was outside this task's scope.
1108
+
1109
  ## Limitations
1110
 
1111
  - No Lean compilation, proof checking, Comparator run or mathematical verification of any kind.
1112
+ - The build ran in a sandbox-free subprocess only for the generator; `cryptids`, `formal-conjectures`,
1113
+ `formal-imo`, `formal-putnam-like`, `alphageometry`, `alphaproof-nexus` and the notebooks were never
1114
+ built or executed, only read as text.
1115
  - PDF prose proofs are noisy pypdf text; symbol and ligature fidelity is not guaranteed.
1116
  - The generator subset is a seeded sample of one curriculum, not the paper's full dataset, and its
1117
+ questions are school-level rather than research mathematics. One module, `algebra__linear_1d`, stopped
1118
+ at 3,055 of its 8,000 quota when upstream's own `assert entropy == 0` fired; nothing was repaired, and
1119
+ the shortfall is recorded in `audit/mathematics-dataset-generation.json`.
1120
  - `formal-conjectures` upstream states that some content derives from third-party sources (Wikipedia,
1121
  MathOverflow and OEIS under CC-BY-SA 4.0). Per-file reference URLs are retained inside the Lean text; no
1122
  separate per-file license audit was performed.
1123
+ - Near-duplicate detection, Phase 1 overlap analysis and editorial review were not performed. 1,629
1124
+ exact duplicate texts were dropped and are listed in `audit/excluded-duplicates.jsonl`.
1125
  - The SentencePiece model normalises runs of spaces, so packed text decodes whitespace-normalised rather
1126
  than byte-identically to the sources.
1127
+ - Only problems that upstream published are present, so several subsets are selection-biased. `cryptids`
1128
+ contributes 31.6M of the 78.4M characters because a handful of Busy Beaver traces are enormous, so the
1129
+ corpus is dominated by machine-generated proof script rather than mathematics prose.
1130
 
1131
  ## Statistics
1132