thai_doc_parsing v1 batch b1

README.md

@@ -10,49 +10,67 @@ configs:
   - config_name: pages
     data_files:
       - split: test
-        path: "pages-test-*.tar"
+        path:
+          - "pages-test-000000.tar"
       - split: train
-        path: "pages-train-*.tar"
+        path:
+          - "pages-train-000000.tar"
+          - "pages-train-000001.tar"
+          - "pages-train-000002.tar"
+          - "pages-train-000003.tar"
+          - "pages-train-000004.tar"
       - split: validation
-        path: "pages-validation-*.tar"
+        path:
+          - "pages-validation-000000.tar"
   - config_name: documents
     data_files:
       - split: test
-        path: "documents-test-*.tar"
+        path:
+          - "documents-test-000000.tar"
       - split: train
-        path: "documents-train-*.tar"
+        path:
+          - "documents-train-000000.tar"
+          - "documents-train-000001.tar"
       - split: validation
-        path: "documents-validation-*.tar"
+        path:
+          - "documents-validation-000000.tar"
 ---
 
 # Thai document parsing corpus (thai_doc_parsing v1)
 
-Structured Thai documents for vision-language continued pretraining: page images, one
-canonical annotation (`thai-doc-ir/1`), and deterministic Markdown,
-LaTeX,
-/ Typst / HTML views.
+Structured Thai documents for vision-language continued pretraining: page images, one canonical
+annotation (`thai-doc-ir/1`), and deterministic Markdown / LaTeX / Typst / HTML views.
 
-## Loading
+## Sources
+Government publications and forms from the data.go.th catalog and allowlisted agency hosts,
+plus the sources recorded in `sources.parquet`. Every document keeps its source URL, landing
+page snapshot hash, retrieval time, and the licence recorded with it.
 
-```python
-import datasets
+## Rights basis
+Only documents whose item-level decision permits redistribution are packaged. Each row of
+`rights.parquet` carries `training_allowed`, `redistribution_allowed`, the licence, the evidence
+URL, the terms-snapshot hash, and the reviewer.
+Laned counts in this release: {"green": 4939, "amber": 1028}.
+Material with unresolved rights stays in local quarantine and is not uploaded.
 
-pages = datasets.load_dataset("smartwhatt/thai_doc_parsing", "pages", split="train")
-documents = datasets.load_dataset("smartwhatt/thai_doc_parsing", "documents", split="train")
-```
+## Exclusions
+Documents with personal data (national identifiers, contact details, signatures, filled forms),
+unknown licences, non-commercial-only licences, or failing quality gates are excluded; the counts
+and reasons are in `release.json`.
 
-| Config | Sample | Fields |
-| --- | --- | --- |
-| `pages` (default) | one annotated page, key `<doc_id>_pNNNNNN` | `png` (page image), `txt` (its text in reading order), `ir` (the `thai-doc-ir/1` annotation as JSON text), `json` (flat metadata) |
-| `documents` | one document, key `<doc_id>` | `md`, `tex`, `typst`, `html` (four views one annotation), `pdf` (the source bytes), `bundle` (metadata, coverage, rights and every page's IR as JSON text), `json` (flat metadata) |
-
-Shards are named `<family>-<split>-NNNNNN.tar`; a page's split is its document's, so a family never crosses a split boundary.
+## Intended use and limitations
+Continued pretraining and evaluation of Thai document understanding. Known limitations: the
+first tranche is government-publication heavy
+({"government": 8920, "academic": 100, "legal": 1} by family),
+historical scans are represented only where triage routed them to the OCR/VLM lanes, and
+complex tables and equations are serialized with declared transformations rather than a
+guaranteed lossless reconstruction.
 
 ## Token counts
-
-Measured with the pinned tokenizer and image processor (`Qwen/Qwen3-VL-2B-Instruct@main`). The four serializations are four views of
-one annotation: they are listed separately and not summed, and only the selected view
-contributes a document's semantic text tokens.
+Measured with the pinned tokenizer and image processor (`Qwen/Qwen3-VL-2B-Instruct@main`), never estimated from
+a pages-to-tokens multiplier. The four serializations are four views of one annotation: they are
+listed separately and **not** summed, and only the selected view contributes a document's semantic
+text tokens.
 
 | Text view | Files | Text tokens |
 | --- | --- | --- |
@@ -67,17 +85,8 @@ contributes a document's semantic text tokens.
 | **Visual (page images, 15,240 pages)** | **57,168,915** |
 | **Combined for one training pass** | **70,943,826** |
 
-## Release
-
-* Documents: 1,819
-* Pages: 15,240
-* Shards: 11 (12.3 GB)
-* Splits: {}
-* Parser revisions: ["Qwen/Qwen3-VL-2B-Instruct@main#prompt:a93e0c2205d2", "pypdfium2-unavailable"]
-* Rights basis: only documents whose item-level decision permits redistribution are packaged; see `rights.parquet` and `attribution.jsonl`
-* Config hash `6c43ba7d42a75b8a`, code commit `a8c5570e6f2699d1fb8546df27018ff4abe9a994`, built 2026-10-04T21:10:27+00:00
 
 ## Provenance and takedown
-
-Every shard maps to document ids through `release.json`; open an issue on this repository for a
-correction or takedown.
+Code commit `c37ba18d3217d44cd73a02d8dca8a366bd6e9778`, config hash `6c43ba7d42a75b8a`, IR version `thai-doc-ir/1`.
+Takedown or correction requests: open an issue on the dataset repository; every shard maps to
+document ids through `release.json`, so an affected document can be located across all artifacts.

checksums.sha256

@@ -1,11 +1,11 @@
-e46c71504723ca4c8b05406e0e827b86255a7e90b1d17cbf2b9cf58e9f817730  README.md
+9a5c406867eaa5685a8503370fd6ec3a1e4c6c82b378c7935d009acf255ac520  README.md
 5692f55340cc30debe37b480cddb27d2512c67856973f23cdae243d8171e7e03  attribution.jsonl
 0bdddb137d894c14ab48840725fb362bb8c64e5987a4e7744ef624b7ba046718  canonical.parquet
-e46c71504723ca4c8b05406e0e827b86255a7e90b1d17cbf2b9cf58e9f817730  dataset_card.md
-772bb083fa36b2b3ca8764866d208eb8491edeed6bdd52fbf3ccaea608759198  documents-test-000000.tar
-1a468a441883f0e97bd20b79727a9cb41a9429aca0bc7360043d39e6f21b179d  documents-train-000000.tar
-1c31b402bb6e9af60445a312be9f964e82e68b4a05c4e54effd21abd782fcf61  documents-train-000001.tar
-2eb44fd801daf9a31059b7052e13123a375b9c8afc675541f86afa6d98e391be  documents-validation-000000.tar
+9a5c406867eaa5685a8503370fd6ec3a1e4c6c82b378c7935d009acf255ac520  dataset_card.md
+e770e6831e16b28ffea832cc0b2d5c8eabbe8d274eb75fb00f1442c100127204  documents-test-000000.tar
+e6fe509180d61b57ab16d2019d4aaf368f4abf66d2b3e302831a258852988685  documents-train-000000.tar
+9e4c08277ad246983688f4b73110b2d9f464398154923ed085f7335442164563  documents-train-000001.tar
+af1b91acf0e4953a93a118010b7e32d87a2b59bd7cccb4faefab6fee1671ef66  documents-validation-000000.tar
 463d8f415a5842fee25bbd9937c76c8f18223690eac437b5a101e496b0054321  documents.parquet
 9512071d14fb0768cbaca27e0e2f6f159a3fdb5a086c511221842fc0639e4506  duplicate_families.parquet
 faffdb0edfa472a29479a215609ff0053962ba73024da4296d68533cc220a3ac  pages-test-000000.tar
@@ -17,5 +17,5 @@ be6009b60b6e211cdd219dccdc297eef06a2f8a0d2fb853f55ec78aebf439f40  pages-train-00
 e1da6b406774cb6db66d62baf9594637a3a6d010073c20105ae1e772a6380777  pages-validation-000000.tar
 de755e54e21d8e402e9bc4079aa6fe3e943947e9bbf0dd1676d8fc4ad30f8744  pages.parquet
 3c9229a9af4f453f10cdac463026ea8a1b9d32961042766d48d590dcf2430350  privacy.parquet
-c8fc9da117bdc8189886129fe3a6c208521c02a53a650b759536e36b7d04b565  release.json
+592afb475616235b55d446eb2a76490e7bf10289ded82a3ef36325ad35557576  release.json
 38b30a5d3f55f916884109a4a6e6b40e90af93b550ec57badbdabe083c666ed3  rights.parquet

dataset_card.md

@@ -10,49 +10,67 @@ configs:
   - config_name: pages
     data_files:
       - split: test
-        path: "pages-test-*.tar"
+        path:
+          - "pages-test-000000.tar"
       - split: train
-        path: "pages-train-*.tar"
+        path:
+          - "pages-train-000000.tar"
+          - "pages-train-000001.tar"
+          - "pages-train-000002.tar"
+          - "pages-train-000003.tar"
+          - "pages-train-000004.tar"
       - split: validation
-        path: "pages-validation-*.tar"
+        path:
+          - "pages-validation-000000.tar"
   - config_name: documents
     data_files:
       - split: test
-        path: "documents-test-*.tar"
+        path:
+          - "documents-test-000000.tar"
       - split: train
-        path: "documents-train-*.tar"
+        path:
+          - "documents-train-000000.tar"
+          - "documents-train-000001.tar"
       - split: validation
-        path: "documents-validation-*.tar"
+        path:
+          - "documents-validation-000000.tar"
 ---
 
 # Thai document parsing corpus (thai_doc_parsing v1)
 
-Structured Thai documents for vision-language continued pretraining: page images, one
-canonical annotation (`thai-doc-ir/1`), and deterministic Markdown,
-LaTeX,
-/ Typst / HTML views.
+Structured Thai documents for vision-language continued pretraining: page images, one canonical
+annotation (`thai-doc-ir/1`), and deterministic Markdown / LaTeX / Typst / HTML views.
 
-## Loading
+## Sources
+Government publications and forms from the data.go.th catalog and allowlisted agency hosts,
+plus the sources recorded in `sources.parquet`. Every document keeps its source URL, landing
+page snapshot hash, retrieval time, and the licence recorded with it.
 
-```python
-import datasets
+## Rights basis
+Only documents whose item-level decision permits redistribution are packaged. Each row of
+`rights.parquet` carries `training_allowed`, `redistribution_allowed`, the licence, the evidence
+URL, the terms-snapshot hash, and the reviewer.
+Laned counts in this release: {"green": 4939, "amber": 1028}.
+Material with unresolved rights stays in local quarantine and is not uploaded.
 
-pages = datasets.load_dataset("smartwhatt/thai_doc_parsing", "pages", split="train")
-documents = datasets.load_dataset("smartwhatt/thai_doc_parsing", "documents", split="train")
-```
+## Exclusions
+Documents with personal data (national identifiers, contact details, signatures, filled forms),
+unknown licences, non-commercial-only licences, or failing quality gates are excluded; the counts
+and reasons are in `release.json`.
 
-| Config | Sample | Fields |
-| --- | --- | --- |
-| `pages` (default) | one annotated page, key `<doc_id>_pNNNNNN` | `png` (page image), `txt` (its text in reading order), `ir` (the `thai-doc-ir/1` annotation as JSON text), `json` (flat metadata) |
-| `documents` | one document, key `<doc_id>` | `md`, `tex`, `typst`, `html` (four views one annotation), `pdf` (the source bytes), `bundle` (metadata, coverage, rights and every page's IR as JSON text), `json` (flat metadata) |
-
-Shards are named `<family>-<split>-NNNNNN.tar`; a page's split is its document's, so a family never crosses a split boundary.
+## Intended use and limitations
+Continued pretraining and evaluation of Thai document understanding. Known limitations: the
+first tranche is government-publication heavy
+({"government": 8920, "academic": 100, "legal": 1} by family),
+historical scans are represented only where triage routed them to the OCR/VLM lanes, and
+complex tables and equations are serialized with declared transformations rather than a
+guaranteed lossless reconstruction.
 
 ## Token counts
-
-Measured with the pinned tokenizer and image processor (`Qwen/Qwen3-VL-2B-Instruct@main`). The four serializations are four views of
-one annotation: they are listed separately and not summed, and only the selected view
-contributes a document's semantic text tokens.
+Measured with the pinned tokenizer and image processor (`Qwen/Qwen3-VL-2B-Instruct@main`), never estimated from
+a pages-to-tokens multiplier. The four serializations are four views of one annotation: they are
+listed separately and **not** summed, and only the selected view contributes a document's semantic
+text tokens.
 
 | Text view | Files | Text tokens |
 | --- | --- | --- |
@@ -67,17 +85,8 @@ contributes a document's semantic text tokens.
 | **Visual (page images, 15,240 pages)** | **57,168,915** |
 | **Combined for one training pass** | **70,943,826** |
 
-## Release
-
-* Documents: 1,819
-* Pages: 15,240
-* Shards: 11 (12.3 GB)
-* Splits: {}
-* Parser revisions: ["Qwen/Qwen3-VL-2B-Instruct@main#prompt:a93e0c2205d2", "pypdfium2-unavailable"]
-* Rights basis: only documents whose item-level decision permits redistribution are packaged; see `rights.parquet` and `attribution.jsonl`
-* Config hash `6c43ba7d42a75b8a`, code commit `a8c5570e6f2699d1fb8546df27018ff4abe9a994`, built 2026-10-04T21:10:27+00:00
 
 ## Provenance and takedown
-
-Every shard maps to document ids through `release.json`; open an issue on this repository for a
-correction or takedown.
+Code commit `c37ba18d3217d44cd73a02d8dca8a366bd6e9778`, config hash `6c43ba7d42a75b8a`, IR version `thai-doc-ir/1`.
+Takedown or correction requests: open an issue on the dataset repository; every shard maps to
+document ids through `release.json`, so an affected document can be located across all artifacts.

documents-test-000000.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:772bb083fa36b2b3ca8764866d208eb8491edeed6bdd52fbf3ccaea608759198
+oid sha256:e770e6831e16b28ffea832cc0b2d5c8eabbe8d274eb75fb00f1442c100127204
 size 179507200

documents-train-000000.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a468a441883f0e97bd20b79727a9cb41a9429aca0bc7360043d39e6f21b179d
+oid sha256:e6fe509180d61b57ab16d2019d4aaf368f4abf66d2b3e302831a258852988685
 size 2006589440

documents-train-000001.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1c31b402bb6e9af60445a312be9f964e82e68b4a05c4e54effd21abd782fcf61
+oid sha256:9e4c08277ad246983688f4b73110b2d9f464398154923ed085f7335442164563
 size 517120000

documents-validation-000000.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2eb44fd801daf9a31059b7052e13123a375b9c8afc675541f86afa6d98e391be
+oid sha256:af1b91acf0e4953a93a118010b7e32d87a2b59bd7cccb4faefab6fee1671ef66
 size 123095040

release.json

@@ -1,8 +1,8 @@
 {
   "batch_id": "b1",
-  "built_at": "2026-10-04T21:10:27+00:00",
+  "built_at": "2026-10-05T05:09:13+00:00",
   "bytes": 12300729235,
-  "code_commit": "a8c5570e6f2699d1fb8546df27018ff4abe9a994",
+  "code_commit": "c37ba18d3217d44cd73a02d8dca8a366bd6e9778",
   "config_hash": "6c43ba7d42a75b8a",
   "dataset": "thai_doc_parsing",
   "document_coverage": {