ZefanCai commited on
Commit
c67699e
·
verified ·
1 Parent(s): b0aad40

Add audited original multilingual mailroom dataset

Browse files

Adds mailroom-control-v1 with 114,800 original typed records, 11,600 email requests and 400 multilingual families in five splits. Includes exact raw JSONL, Parquet, original manifest, auxiliary cases, independent visible-text audit, expansion preservation and pinned source links. No private emails or provider responses are imported. No training or full-corpus model evaluation is claimed; a separate 87-request Jev probe is documented. Existing payloads, configs, manifests, default and frozen mixtures are preserved, including the preceding IR addition.

README.md CHANGED
@@ -3,6 +3,7 @@ license: cc0-1.0
3
  language:
4
  - en
5
  - zh
 
6
  task_categories:
7
  - text-classification
8
  tags:
@@ -147,11 +148,23 @@ configs:
147
  path: data/ir-control-v1/test-*.parquet
148
  - split: ood
149
  path: data/ir-control-v1/ood-*.parquet
 
 
 
 
 
 
 
 
 
 
 
 
150
  ---
151
 
152
  # Open-Jev: typed decision datasets
153
 
154
- Open-Jev turns a state and a question into a typed decision: a yes/no probability, a distribution over choices, independent label probabilities, or a discrete numeric/ordinal decision. This repository publishes eleven separate, frozen data configs from the [Open-Jev project](https://github.com/Zefan-Cai/Open-Jev-Dev), together with original manifests, exact raw records, source code and reconstruction instructions.
155
 
156
  These are controlled, mostly synthetic tasks and reference labels. They are not official TypeSafe/Jev training data, model predictions, or evidence of general capability. Open-Jev is independently implemented and is not affiliated with TypeSafe.
157
 
@@ -170,6 +183,7 @@ These are controlled, mostly synthetic tasks and reference labels. They are not
170
  | `sponsor-segment-control-v1` | 6,345 | 756 | 540 | 999 | 2,160 | 10,800 |
171
  | `silent-failure-control-v1` | 6,432 | 264 | 360 | 624 | 1,920 | 9,600 |
172
  | `ir-control-v1` | 7,366 | 464 | 464 | 986 | 2,320 | 11,600 |
 
173
 
174
  **The configs overlap.** `release-v2-redistributable` is contained in `browser-drone-expansion-v1-redistributable`; do not add config totals and call them unique examples. Counts are typed decision rows. Several heads may come from the same conversation, document, family or game trajectory.
175
 
@@ -203,7 +217,7 @@ For reproducibility, pass `revision="<dataset commit SHA>"`. Choose a config exp
203
  | `record_json` | Complete original JSON record, retaining original object key order and numeric representation. |
204
  | `original_line_number` | One-based row position in the original frozen split, including positions of excluded rows. |
205
 
206
- The Parquet representation avoids imposing one nested schema on different task states. `raw/<config>/<split>.jsonl.gz` preserves the source JSONL bytes after decompression. For the filtered composites, retained lines preserve exact bytes and order. For the nine stand-alone control configs, decompressed files match the original frozen split hashes exactly.
207
 
208
  Use only `state`, `question`, `kind` and `options` as model inputs. Do not expose `target`, `metadata`, identities, split assignments or provenance fields to the model. Distribution, binary, multilabel and ordinal targets have different semantics; do not reduce every row to a single-class accuracy calculation.
209
 
@@ -255,7 +269,7 @@ The new original cases remain separate audit artifacts. Exact raw JSONL bytes, o
255
 
256
  [`ir-control-v1`](cards/ir-control-v1.md) adds **11,600 typed rows across five
257
  splits**, representing 8,800 requests, 400 queries and 200 complete fictional
258
- system families. The repository now declares **11 configs and 55 splits**.
259
  It supports pointwise Noul/Score, pairwise/setwise Choice and listwise
260
  Choice/Score task forms under an original four-level relevance rubric.
261
  Choice labels supervise best-passage selection, not full ranking or uncertainty.
@@ -269,3 +283,26 @@ test/OOD families remain held out in the expanded corpus.
269
  All existing dataset payloads, original manifests, config definitions, the default
270
  config and frozen model-training mixtures remain unchanged. The
271
  [IR export manifest](exports/ir-data-20260920.json) binds this addition only.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  language:
4
  - en
5
  - zh
6
+ - tr
7
  task_categories:
8
  - text-classification
9
  tags:
 
148
  path: data/ir-control-v1/test-*.parquet
149
  - split: ood
150
  path: data/ir-control-v1/ood-*.parquet
151
+ - config_name: mailroom-control-v1
152
+ data_files:
153
+ - split: train
154
+ path: data/mailroom-control-v1/train-*.parquet
155
+ - split: calibration
156
+ path: data/mailroom-control-v1/calibration-*.parquet
157
+ - split: validation
158
+ path: data/mailroom-control-v1/validation-*.parquet
159
+ - split: test
160
+ path: data/mailroom-control-v1/test-*.parquet
161
+ - split: ood
162
+ path: data/mailroom-control-v1/ood-*.parquet
163
  ---
164
 
165
  # Open-Jev: typed decision datasets
166
 
167
+ Open-Jev turns a state and a question into a typed decision: a yes/no probability, a distribution over choices, independent label probabilities, or a discrete numeric/ordinal decision. This repository publishes twelve separate, frozen data configs from the [Open-Jev project](https://github.com/Zefan-Cai/Open-Jev-Dev), together with original manifests, exact raw records, source code and reconstruction instructions.
168
 
169
  These are controlled, mostly synthetic tasks and reference labels. They are not official TypeSafe/Jev training data, model predictions, or evidence of general capability. Open-Jev is independently implemented and is not affiliated with TypeSafe.
170
 
 
183
  | `sponsor-segment-control-v1` | 6,345 | 756 | 540 | 999 | 2,160 | 10,800 |
184
  | `silent-failure-control-v1` | 6,432 | 264 | 360 | 624 | 1,920 | 9,600 |
185
  | `ir-control-v1` | 7,366 | 464 | 464 | 986 | 2,320 | 11,600 |
186
+ | `mailroom-control-v1` | 73,472 | 4,018 | 6,027 | 8,323 | 22,960 | 114,800 |
187
 
188
  **The configs overlap.** `release-v2-redistributable` is contained in `browser-drone-expansion-v1-redistributable`; do not add config totals and call them unique examples. Counts are typed decision rows. Several heads may come from the same conversation, document, family or game trajectory.
189
 
 
217
  | `record_json` | Complete original JSON record, retaining original object key order and numeric representation. |
218
  | `original_line_number` | One-based row position in the original frozen split, including positions of excluded rows. |
219
 
220
+ The Parquet representation avoids imposing one nested schema on different task states. `raw/<config>/<split>.jsonl.gz` preserves the source JSONL bytes after decompression. For the filtered composites, retained lines preserve exact bytes and order. For the ten stand-alone control configs, decompressed files match the original frozen split hashes exactly.
221
 
222
  Use only `state`, `question`, `kind` and `options` as model inputs. Do not expose `target`, `metadata`, identities, split assignments or provenance fields to the model. Distribution, binary, multilabel and ordinal targets have different semantics; do not reduce every row to a single-class accuracy calculation.
223
 
 
269
 
270
  [`ir-control-v1`](cards/ir-control-v1.md) adds **11,600 typed rows across five
271
  splits**, representing 8,800 requests, 400 queries and 200 complete fictional
272
+ system families. That addition brought the repository to **11 configs and 55 splits**.
273
  It supports pointwise Noul/Score, pairwise/setwise Choice and listwise
274
  Choice/Score task forms under an original four-level relevance rubric.
275
  Choice labels supervise best-passage selection, not full ranking or uncertainty.
 
283
  All existing dataset payloads, original manifests, config definitions, the default
284
  config and frozen model-training mixtures remain unchanged. The
285
  [IR export manifest](exports/ir-data-20260920.json) binds this addition only.
286
+
287
+
288
+ ## Original multilingual mailroom addition — 2026-09-20
289
+
290
+ [`mailroom-control-v1`](cards/mailroom-control-v1.md) adds **114,800 typed rows
291
+ across five splits**, from 11,600 original email requests and 400 complete
292
+ English/Chinese/Turkish families. The repository now declares **12 configs and
293
+ 60 splits**. Each request retains the source-shaped two Choice and nine Noul
294
+ questions. Nonbill/nonreceipt categories have no supervised label, and unchanged
295
+ heads across taxonomy variants share their original row. There are 114,800
296
+ unique ordered inputs, or 114,400 after also disregarding candidate order.
297
+
298
+ The independent auditor derives labels from final visible email text and
299
+ taxonomy. All frozen probe rows/cases keep their bytes and split assignments.
300
+ No real mailbox contents, private source examples, attachments or provider
301
+ responses are included. No training or full-corpus model evaluation is claimed;
302
+ the separately documented 87-request Jev probe is not production mail-triage
303
+ accuracy or end-to-end mailbox automation.
304
+
305
+ All existing payloads, manifests, config definitions, the default config and
306
+ frozen training mixtures remain unchanged, including the preceding IR addition.
307
+ The [mailroom export manifest](exports/mailroom-data-20260920.json) binds only
308
+ this new config and its auxiliary audit artifacts.
artifacts/mailroom-control-v1/cases.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7cda8e04e7bfc884d1ffba9e920c4108e45d80198e65952d3d1d4e93282f6b9
3
+ size 2237258
cards/mailroom-control-v1.md ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Original multilingual mailroom controls
2
+
3
+ **114,800 typed rows, 11,600 email requests and 400 whole synthetic families.**
4
+ Original English, Chinese and Turkish emails and labels are CC0-1.0. Source
5
+ code is MIT. No real mailbox, source emails, IMAP contents, attachments or the
6
+ community author's private examples were imported.
7
+
8
+ The [pinned community implementation](https://github.com/selcukusta/jev-mailroom/tree/06d44889231afb209e29275f14a529ba52c9eb0d)
9
+ asks two Choice and nine Noul questions, eleven total. Our implementation
10
+ preserves state shape `email.subject`, `email.from`, `email.date`, `email.body`
11
+ and the eleven-head contract with independently authored question wording.
12
+
13
+ `kind` distinguishes invoice, payment confirmation, account statement,
14
+ promotion, newsletter and other. Four evidence Nouls ask about currently owed
15
+ money, the sender issuing its own charges, billing identifiers and promotion.
16
+ `category` distinguishes education, electricity, telecom, banking, airline and
17
+ other; five independent Nouls also ask about these subjects. Category is
18
+ unsupervised for nonbill/nonreceipt messages: 4,800 requests retain the runtime
19
+ question but omit its training label. An application fallback is not ground truth.
20
+
21
+ Already-paid amounts and advertised prices are not currently owed money.
22
+ Invoice links without repeated amounts remain invoices; relaying a bill does
23
+ not make the relay its issuer. Controls distinguish educational services from
24
+ a student's phone bill, handset retail from telecom service, equipment from
25
+ electricity supply, bank-mediated payment from a bank's own charges, and known
26
+ out-of-taxonomy services. Unknown or contradictory grammar is not assigned gold.
27
+
28
+ Each family contains three languages and nine actions, plus two English invoice
29
+ taxonomy variants. Reordering categories tests order sensitivity; adding
30
+ insurance changes the insurance label only when that option exists. No change
31
+ in confidence is assumed or supervised. Shared unchanged question inputs across
32
+ taxonomy variants reference their first row. There are **114,800 unique ordered
33
+ inputs**, or **114,400** when candidate order is also disregarded.
34
+
35
+ | Split | Typed rows |
36
+ | --- | ---: |
37
+ | train | 73,472 |
38
+ | calibration | 4,018 |
39
+ | validation | 6,027 |
40
+ | test | 8,323 |
41
+ | ood | 22,960 |
42
+
43
+ All translations, actions, services and taxonomies for a family share a split.
44
+ Index modulo ten values 8–9 reserve OOD; other families use stable hashing.
45
+ OOD reserves fictional identities and moves body paragraphs when possible;
46
+ two-paragraph messages retain their order. It is not a new grammar for every
47
+ request. The frozen 10-family probe stays byte-identical in the expanded corpus,
48
+ including all 290 requests and 2,870 rows in their original split files.
49
+
50
+ Use only state, question, kind and options as model inputs. Targets, metadata,
51
+ case reference labels and masks are privileged audit fields. Each Parquet
52
+ `record_json` and decompressed raw split retain exact original JSONL bytes and
53
+ row order, with the same column schema as the existing configs.
54
+
55
+ **No training on this corpus is claimed.** A separate real Jev 1.13.0 probe
56
+ completed 87 held-out requests containing 957 runtime questions and 921 labelled
57
+ decisions; 36 category questions had no gold. Its 908/921 correct decisions are
58
+ a small correlated control probe, not full-corpus evaluation or production
59
+ multilingual accuracy. No provider responses are included in this data upload.
60
+ The generation manifest records operations at generation time; later probe
61
+ evaluation is documented separately. Frozen model-training mixtures are unchanged.
62
+
63
+ - [Original manifest](../provenance/original-manifests/mailroom-control-v1.json)
64
+ - [Independent body audit](../provenance/mailroom-data-20260920/independent-audit.json)
65
+ - [Expansion preservation](../provenance/mailroom-data-20260920/expansion-stability.json)
66
+ - [Auxiliary request cases](../artifacts/mailroom-control-v1/cases.jsonl.gz)
67
+ - [Pinned generator](https://github.com/Zefan-Cai/Open-Jev/blob/61fc74408e95fa1844c1bd118245c2c871a84585/jev/mailroom_data.py), [auditor](https://github.com/Zefan-Cai/Open-Jev/blob/61fc74408e95fa1844c1bd118245c2c871a84585/jev/mailroom_audit.py) and [documentation](https://github.com/Zefan-Cai/Open-Jev/blob/61fc74408e95fa1844c1bd118245c2c871a84585/docs/mailroom-control-data.md)
68
+ - [Incremental export manifest](../exports/mailroom-data-20260920.json)
69
+
70
+ From the pinned source checkout, choose a fresh output directory:
71
+
72
+ ```bash
73
+ python3 -m jev.mailroom_data --output-dir data/mailroom-control-v1-rebuilt --groups 400 --seed 943
74
+ python3 -m jev.mailroom_audit --data data/mailroom-control-v1-rebuilt --output mailroom-audit.json
75
+ ```
76
+
77
+ This is a finite original-data task, not a collection of natural private mail,
78
+ live mailbox automation, or an independently annotated production benchmark.
data/mailroom-control-v1/calibration-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad584458b9d498702d2c3fd5cb4427a8ea57e588ad3cc31a3cca3609213ccc3b
3
+ size 376113
data/mailroom-control-v1/ood-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d14896ef69f5ac52899d93cd019f80409373622f7707346c4c9fac9814a3964
3
+ size 2155618
data/mailroom-control-v1/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:605a9f4da39da089f8f9020e57ec57999ed5451ff3506c5629dbd056d5594280
3
+ size 795314
data/mailroom-control-v1/train-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5bc38433ab3758fc31d56a2f4f9e1aebb58881d1f21edf1bffdce49f38d43d32
3
+ size 6882318
data/mailroom-control-v1/validation-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c60276b58d5856173b316266c5323c6683049ccca70b479210871371632ca317
3
+ size 565833
exports/mailroom-data-20260920.json ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "repo_id": "ZefanCai/Open-Jev",
4
+ "release": "mailroom-data-20260920",
5
+ "parent_dataset_revision": "b0aad4004b8d74f4a6ca66c7a9175fe17478d688",
6
+ "operation": "Add mailroom-control-v1 five splits and exact raw/auxiliary records, manifests, audit and card; update root card only",
7
+ "new_data_training_performed": false,
8
+ "full_corpus_model_evaluation_performed": false,
9
+ "separate_frozen_probe_evaluation": {
10
+ "provider": "Jev 1.13.0",
11
+ "requests": 87,
12
+ "runtime_questions": 957,
13
+ "labelled_decisions": 921,
14
+ "correct": 908,
15
+ "raw_model_responses_included": false
16
+ },
17
+ "real_mail_imported": false,
18
+ "frozen_training_mixture_changed": false,
19
+ "frozen_pilot_manifest_sha256": "6ec887c96072c7bf862a82fef5f65cc651a2cd7f0ac59f240b3cb9f0c849c572",
20
+ "source_commit": "61fc74408e95fa1844c1bd118245c2c871a84585",
21
+ "source_blob_base_url": "https://github.com/Zefan-Cai/Open-Jev/blob/61fc74408e95fa1844c1bd118245c2c871a84585",
22
+ "source_code_sha256": {
23
+ "jev/mailroom_data.py": "cd84ef39473153ec23cfebb426db9810ea68b3fcda8040d761671ae282af1a49",
24
+ "jev/mailroom_audit.py": "20bdd9de36d06f0591889c85f4a634f9f748e8a15125fd686eae26dcf8c09e09",
25
+ "jev/mailroom_probe.py": "621a2fb6e9c95e9bff69bef8f73dc3a6cf053968bf33b3d8795c72c92ad290a4"
26
+ },
27
+ "configs": {
28
+ "mailroom-control-v1": {
29
+ "total": 114800,
30
+ "groups": 400,
31
+ "request_case_count": 11600,
32
+ "ordered_unique_model_inputs": 114800,
33
+ "candidate_order_invariant_unique_inputs": 114400,
34
+ "masked_category_cases": 4800,
35
+ "license": "CC0-1.0",
36
+ "original_manifest_sha256": "afa9daabd5639a333c9adf8e89e7c7bbfcf84adfc051fb9843978c29703122f6",
37
+ "splits": {
38
+ "train": {
39
+ "count": 73472,
40
+ "original_sha256": "5a19ebf41125093c8bb7aa0861842673686014b2c288e79bfbb54a0ea1a4a86d",
41
+ "raw_path": "raw/mailroom-control-v1/train.jsonl.gz",
42
+ "raw_uncompressed_sha256": "5a19ebf41125093c8bb7aa0861842673686014b2c288e79bfbb54a0ea1a4a86d",
43
+ "raw_sha256": "ebbfae0b4024ac8dbbbc709f54b8cc02023d5953a0b712888a9d2ffbb5189811",
44
+ "parquet_path": "data/mailroom-control-v1/train-00000-of-00001.parquet",
45
+ "parquet_sha256": "5bc38433ab3758fc31d56a2f4f9e1aebb58881d1f21edf1bffdce49f38d43d32",
46
+ "original_line_numbers_preserved": true,
47
+ "sources": {
48
+ "mailroom-control-v1": 73472
49
+ },
50
+ "kinds": {
51
+ "choice": 11264,
52
+ "noul": 62208
53
+ }
54
+ },
55
+ "calibration": {
56
+ "count": 4018,
57
+ "original_sha256": "9765f7481ff6eda49733d1cdee648c771bb1a2055bcc7df9bac5d2acda9b592b",
58
+ "raw_path": "raw/mailroom-control-v1/calibration.jsonl.gz",
59
+ "raw_uncompressed_sha256": "9765f7481ff6eda49733d1cdee648c771bb1a2055bcc7df9bac5d2acda9b592b",
60
+ "raw_sha256": "8cd86ca00c6d75b3ee356e04d8ca872f6442e5a0566e5a50c9af18709a6a130f",
61
+ "parquet_path": "data/mailroom-control-v1/calibration-00000-of-00001.parquet",
62
+ "parquet_sha256": "ad584458b9d498702d2c3fd5cb4427a8ea57e588ad3cc31a3cca3609213ccc3b",
63
+ "original_line_numbers_preserved": true,
64
+ "sources": {
65
+ "mailroom-control-v1": 4018
66
+ },
67
+ "kinds": {
68
+ "choice": 616,
69
+ "noul": 3402
70
+ }
71
+ },
72
+ "validation": {
73
+ "count": 6027,
74
+ "original_sha256": "97558afbd86cb84cada925c9513fa868db3e47db83e51fc1dafe955240d05516",
75
+ "raw_path": "raw/mailroom-control-v1/validation.jsonl.gz",
76
+ "raw_uncompressed_sha256": "97558afbd86cb84cada925c9513fa868db3e47db83e51fc1dafe955240d05516",
77
+ "raw_sha256": "5705a3babf6b45a231946e48a4a2eed41c81e8d5ec9de1de42ec69801da34db4",
78
+ "parquet_path": "data/mailroom-control-v1/validation-00000-of-00001.parquet",
79
+ "parquet_sha256": "c60276b58d5856173b316266c5323c6683049ccca70b479210871371632ca317",
80
+ "original_line_numbers_preserved": true,
81
+ "sources": {
82
+ "mailroom-control-v1": 6027
83
+ },
84
+ "kinds": {
85
+ "choice": 924,
86
+ "noul": 5103
87
+ }
88
+ },
89
+ "test": {
90
+ "count": 8323,
91
+ "original_sha256": "9bb9976cfa9e3e5dc084b86d9cbeb0bc201465a5ca0249af6781c02bdf635938",
92
+ "raw_path": "raw/mailroom-control-v1/test.jsonl.gz",
93
+ "raw_uncompressed_sha256": "9bb9976cfa9e3e5dc084b86d9cbeb0bc201465a5ca0249af6781c02bdf635938",
94
+ "raw_sha256": "d6662f7f7870d3bc629956f07fed8dfb5f0a42d1a923fbbe764db30e933036be",
95
+ "parquet_path": "data/mailroom-control-v1/test-00000-of-00001.parquet",
96
+ "parquet_sha256": "605a9f4da39da089f8f9020e57ec57999ed5451ff3506c5629dbd056d5594280",
97
+ "original_line_numbers_preserved": true,
98
+ "sources": {
99
+ "mailroom-control-v1": 8323
100
+ },
101
+ "kinds": {
102
+ "choice": 1276,
103
+ "noul": 7047
104
+ }
105
+ },
106
+ "ood": {
107
+ "count": 22960,
108
+ "original_sha256": "5667325039486c28ade0b752c3fbf694aa7c73e7998a274744b09b5dea207d0e",
109
+ "raw_path": "raw/mailroom-control-v1/ood.jsonl.gz",
110
+ "raw_uncompressed_sha256": "5667325039486c28ade0b752c3fbf694aa7c73e7998a274744b09b5dea207d0e",
111
+ "raw_sha256": "e63f4949c9103b26a718e7df2a4416161f46093e27569a34de64083c587dd941",
112
+ "parquet_path": "data/mailroom-control-v1/ood-00000-of-00001.parquet",
113
+ "parquet_sha256": "4d14896ef69f5ac52899d93cd019f80409373622f7707346c4c9fac9814a3964",
114
+ "original_line_numbers_preserved": true,
115
+ "sources": {
116
+ "mailroom-control-v1": 22960
117
+ },
118
+ "kinds": {
119
+ "choice": 3520,
120
+ "noul": 19440
121
+ }
122
+ }
123
+ },
124
+ "extra_artifacts": {
125
+ "artifacts/mailroom-control-v1/cases.jsonl.gz": {
126
+ "uncompressed_sha256": "5872ce885ff3f3eac6cf4a605db25d3c176fcfce10fbc9aefd28bb601e269c2d",
127
+ "uncompressed_bytes": 68978600,
128
+ "sha256": "f7cda8e04e7bfc884d1ffba9e920c4108e45d80198e65952d3d1d4e93282f6b9",
129
+ "bytes": 2237258
130
+ }
131
+ }
132
+ }
133
+ },
134
+ "published_files": {
135
+ "README.md": {
136
+ "sha256": "8f7d66430c63881d042a88165f4ad669e6b7c6783b69f6735b8dfedf6d58f683",
137
+ "bytes": 20866
138
+ },
139
+ "artifacts/mailroom-control-v1/cases.jsonl.gz": {
140
+ "sha256": "f7cda8e04e7bfc884d1ffba9e920c4108e45d80198e65952d3d1d4e93282f6b9",
141
+ "bytes": 2237258
142
+ },
143
+ "cards/mailroom-control-v1.md": {
144
+ "sha256": "998ad29c16c38306a384560c4bd364dc74ace43be3428f5527e887ef8dc56557",
145
+ "bytes": 4828
146
+ },
147
+ "data/mailroom-control-v1/calibration-00000-of-00001.parquet": {
148
+ "sha256": "ad584458b9d498702d2c3fd5cb4427a8ea57e588ad3cc31a3cca3609213ccc3b",
149
+ "bytes": 376113
150
+ },
151
+ "data/mailroom-control-v1/ood-00000-of-00001.parquet": {
152
+ "sha256": "4d14896ef69f5ac52899d93cd019f80409373622f7707346c4c9fac9814a3964",
153
+ "bytes": 2155618
154
+ },
155
+ "data/mailroom-control-v1/test-00000-of-00001.parquet": {
156
+ "sha256": "605a9f4da39da089f8f9020e57ec57999ed5451ff3506c5629dbd056d5594280",
157
+ "bytes": 795314
158
+ },
159
+ "data/mailroom-control-v1/train-00000-of-00001.parquet": {
160
+ "sha256": "5bc38433ab3758fc31d56a2f4f9e1aebb58881d1f21edf1bffdce49f38d43d32",
161
+ "bytes": 6882318
162
+ },
163
+ "data/mailroom-control-v1/validation-00000-of-00001.parquet": {
164
+ "sha256": "c60276b58d5856173b316266c5323c6683049ccca70b479210871371632ca317",
165
+ "bytes": 565833
166
+ },
167
+ "provenance/mailroom-data-20260920/expansion-stability.json": {
168
+ "sha256": "0537cb437726ac4cf2422dead14d5bc4678fbc53aa773f4d91358902af6b6a6e",
169
+ "bytes": 906
170
+ },
171
+ "provenance/mailroom-data-20260920/independent-audit.json": {
172
+ "sha256": "e177a405931c50b3852902dd7c7aa622717a65f3788dba21eed47d5a1b04f296",
173
+ "bytes": 1758
174
+ },
175
+ "provenance/original-manifests/mailroom-control-v1.json": {
176
+ "sha256": "afa9daabd5639a333c9adf8e89e7c7bbfcf84adfc051fb9843978c29703122f6",
177
+ "bytes": 2756
178
+ },
179
+ "raw/mailroom-control-v1/calibration.jsonl.gz": {
180
+ "sha256": "8cd86ca00c6d75b3ee356e04d8ca872f6442e5a0566e5a50c9af18709a6a130f",
181
+ "bytes": 130633
182
+ },
183
+ "raw/mailroom-control-v1/ood.jsonl.gz": {
184
+ "sha256": "e63f4949c9103b26a718e7df2a4416161f46093e27569a34de64083c587dd941",
185
+ "bytes": 739255
186
+ },
187
+ "raw/mailroom-control-v1/test.jsonl.gz": {
188
+ "sha256": "d6662f7f7870d3bc629956f07fed8dfb5f0a42d1a923fbbe764db30e933036be",
189
+ "bytes": 268265
190
+ },
191
+ "raw/mailroom-control-v1/train.jsonl.gz": {
192
+ "sha256": "ebbfae0b4024ac8dbbbc709f54b8cc02023d5953a0b712888a9d2ffbb5189811",
193
+ "bytes": 2350124
194
+ },
195
+ "raw/mailroom-control-v1/validation.jsonl.gz": {
196
+ "sha256": "5705a3babf6b45a231946e48a4a2eed41c81e8d5ec9de1de42ec69801da34db4",
197
+ "bytes": 195258
198
+ }
199
+ }
200
+ }
provenance/mailroom-data-20260920/expansion-stability.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "verified": true,
3
+ "small_groups": 10,
4
+ "large_groups": 400,
5
+ "preserved_in_same_files": {
6
+ "train.jsonl": 2009,
7
+ "calibration.jsonl": 0,
8
+ "validation.jsonl": 0,
9
+ "test.jsonl": 287,
10
+ "ood.jsonl": 574,
11
+ "cases.jsonl": 290
12
+ },
13
+ "small_manifest_sha256": "6ec887c96072c7bf862a82fef5f65cc651a2cd7f0ac59f240b3cb9f0c849c572",
14
+ "large_manifest_sha256": "afa9daabd5639a333c9adf8e89e7c7bbfcf84adfc051fb9843978c29703122f6",
15
+ "generator_sha256": "cd84ef39473153ec23cfebb426db9810ea68b3fcda8040d761671ae282af1a49",
16
+ "meaning": "Every small-corpus JSONL line occurs byte-identically, exactly once, in the same full-corpus file. Thus test/OOD requests and rows never move to train. Full-corpus additional holdout families do not alter the frozen probe selection.",
17
+ "model_outputs_consulted": false,
18
+ "verifier_sha256": "65156fa741acbe4fd52da7c10bbddb3d3c39a90c17e0082db73eb6fea853d73a"
19
+ }
provenance/mailroom-data-20260920/independent-audit.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "verified": true,
3
+ "cases": 11600,
4
+ "records": 114800,
5
+ "groups": 400,
6
+ "group_splits": {
7
+ "train": 256,
8
+ "test": 29,
9
+ "ood": 80,
10
+ "calibration": 14,
11
+ "validation": 21
12
+ },
13
+ "languages": {
14
+ "en": 4400,
15
+ "zh": 3600,
16
+ "tr": 3600
17
+ },
18
+ "masked_questions": {
19
+ "category": 4800
20
+ },
21
+ "label_counts": {
22
+ "about_airline:False": 10680,
23
+ "about_airline:True": 920,
24
+ "about_banking:False": 9480,
25
+ "about_banking:True": 2120,
26
+ "about_education:False": 10680,
27
+ "about_education:True": 920,
28
+ "about_electricity:False": 10680,
29
+ "about_electricity:True": 920,
30
+ "about_telecom:False": 10680,
31
+ "about_telecom:True": 920,
32
+ "category:airline": 680,
33
+ "category:banking": 680,
34
+ "category:education": 680,
35
+ "category:electricity": 680,
36
+ "category:insurance": 40,
37
+ "category:other": 3360,
38
+ "category:telecom": 680,
39
+ "from_billing_entity:False": 7200,
40
+ "from_billing_entity:True": 4400,
41
+ "has_billing_identifiers:False": 4800,
42
+ "has_billing_identifiers:True": 6800,
43
+ "is_promotional:False": 10400,
44
+ "is_promotional:True": 1200,
45
+ "kind:account_statement": 1200,
46
+ "kind:invoice": 5600,
47
+ "kind:newsletter": 1200,
48
+ "kind:other": 1200,
49
+ "kind:payment_confirmation": 1200,
50
+ "kind:promotion": 1200,
51
+ "states_amount_owed:False": 7200,
52
+ "states_amount_owed:True": 4400
53
+ },
54
+ "unique_model_inputs": 114800,
55
+ "model_inference_performed": false,
56
+ "scope": "Independent finite-grammar audit from final visible body, sender and taxonomy; no claim of unrestricted semantic labeling.",
57
+ "manifest_sha256": "afa9daabd5639a333c9adf8e89e7c7bbfcf84adfc051fb9843978c29703122f6",
58
+ "auditor_sha256": "20bdd9de36d06f0591889c85f4a634f9f748e8a15125fd686eae26dcf8c09e09"
59
+ }
provenance/original-manifests/mailroom-control-v1.json ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "configuration": {
4
+ "type": "synthetic",
5
+ "generator_version": "mailroom-control-v1",
6
+ "groups": 400,
7
+ "seed": 943,
8
+ "license": "CC0-1.0",
9
+ "source_interface": "https://github.com/selcukusta/jev-mailroom/tree/06d44889231afb209e29275f14a529ba52c9eb0d",
10
+ "source_code_or_emails_imported": false,
11
+ "source_questions_sha256": "d9621c8ccdc9c8bf7dd8132a501390b3d15729a4ef8b1a821ae24d4c14bc5d1d",
12
+ "source_triage_sha256": "b28e64711d2ebc53244fd1d85989158934e1fe0e6b1710294e40a60a3e259cf9",
13
+ "generator_sha256": "cd84ef39473153ec23cfebb426db9810ea68b3fcda8040d761671ae282af1a49",
14
+ "split_policy": "Whole family includes translations, provider/payment/delivery counterfactuals and taxonomy variants. OOD iff index modulo ten >= 8, otherwise stable hash. Expansion preserves every earlier family.",
15
+ "ood_scope": "Unseen fictional families with reordered body paragraphs; finite controlled grammar, not unrestricted natural mail."
16
+ },
17
+ "summary": {
18
+ "records": 114800,
19
+ "groups": 400,
20
+ "unique_inputs": 114400,
21
+ "splits": {
22
+ "calibration": 4018,
23
+ "ood": 22960,
24
+ "test": 8323,
25
+ "train": 73472,
26
+ "validation": 6027
27
+ },
28
+ "kinds": {
29
+ "choice": 17600,
30
+ "noul": 97200
31
+ },
32
+ "families": {
33
+ "evidence": 114800
34
+ },
35
+ "sources": {
36
+ "mailroom-control-v1": 114800
37
+ }
38
+ },
39
+ "files_sha256": {
40
+ "train.jsonl": "5a19ebf41125093c8bb7aa0861842673686014b2c288e79bfbb54a0ea1a4a86d",
41
+ "calibration.jsonl": "9765f7481ff6eda49733d1cdee648c771bb1a2055bcc7df9bac5d2acda9b592b",
42
+ "validation.jsonl": "97558afbd86cb84cada925c9513fa868db3e47db83e51fc1dafe955240d05516",
43
+ "test.jsonl": "9bb9976cfa9e3e5dc084b86d9cbeb0bc201465a5ca0249af6781c02bdf635938",
44
+ "ood.jsonl": "5667325039486c28ade0b752c3fbf694aa7c73e7998a274744b09b5dea207d0e",
45
+ "cases.jsonl": "5872ce885ff3f3eac6cf4a605db25d3c176fcfce10fbc9aefd28bb601e269c2d"
46
+ },
47
+ "model_input_fields": [
48
+ "state",
49
+ "question",
50
+ "kind",
51
+ "options"
52
+ ],
53
+ "request_cases": 11600,
54
+ "language_counts": {
55
+ "en": 4400,
56
+ "zh": 3600,
57
+ "tr": 3600
58
+ },
59
+ "masked_category_cases": 4800,
60
+ "questions_per_request": 11,
61
+ "choice_questions": 2,
62
+ "noul_questions": 9,
63
+ "model_inference_performed": false,
64
+ "training_performed": false,
65
+ "real_mail_imported": false,
66
+ "supervision": "Deterministic semantic labels from original controlled emails. Nonbill category omitted; no confidence, fallback, offline stub, or probability calibration labels.",
67
+ "deduplication": "Identical model inputs across taxonomy variants share the first row id; every request still contains all eleven questions. Case record_ids may reference a row emitted by an earlier case in the same family."
68
+ }
raw/mailroom-control-v1/calibration.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8cd86ca00c6d75b3ee356e04d8ca872f6442e5a0566e5a50c9af18709a6a130f
3
+ size 130633
raw/mailroom-control-v1/ood.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e63f4949c9103b26a718e7df2a4416161f46093e27569a34de64083c587dd941
3
+ size 739255
raw/mailroom-control-v1/test.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6662f7f7870d3bc629956f07fed8dfb5f0a42d1a923fbbe764db30e933036be
3
+ size 268265
raw/mailroom-control-v1/train.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ebbfae0b4024ac8dbbbc709f54b8cc02023d5953a0b712888a9d2ffbb5189811
3
+ size 2350124
raw/mailroom-control-v1/validation.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5705a3babf6b45a231946e48a4a2eed41c81e8d5ec9de1de42ec69801da34db4
3
+ size 195258