Datasets:

Hibiki12y
/

Haebimsource

---
license: other
license_name: source-rights-retained
pretty_name: Haebim Maritime Filtered Corpus
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
- config_name: engineering
  data_files:
  - split: train
    path: engineering/train-*.parquet
- config_name: papers
  data_files:
  - split: train
    path: papers/train-*.parquet
- config_name: tokenized_default
  data_files:
  - split: train
    path: tokenized/default/train-*.parquet
- config_name: tokenized_engineering
  data_files:
  - split: train
    path: tokenized/engineering/train-*.parquet
- config_name: tokenized_papers
  data_files:
  - split: train
    path: tokenized/papers/train-*.parquet
---
# Haebim maritime filtered corpus

Parquet (zstd), converted on 2026-09-30 from the JSONL shards uploaded directly from
PostgreSQL using the Hugging Face API; rows, order and values are unchanged.
Selection: maritime v2 score >= 0.4 (>= 0.655 until 2026-09-30; filter
`maritime_score >= 0.655` to reproduce it), 1024-token runs recorded in manifest.json.
Unscored/rejected and legacy 384-token results are excluded. Full decoded stored
text is included without repair or truncation; scores used at most 4096 characters
and 1024 tokens. Each row preserves URL, source, crawl time, score time, model,
bundle digest, run ID, backend and text SHA256. Cross-URL duplicates are retained.
Scores are machine predictions; the cutoff is not human-gold calibrated.
Source content retains its original rights; no blanket license is assigned.
The original snapshot is preserved. Completed filtering windows were appended at score >= 0.4 until collection stopped on 2026-09-30. Document IDs are deduplicated across original and incremental uploads. Reference-only papers awaiting final review are published separately in the `papers` configuration. The train split is a
corpus partition, not a held-out evaluation set.

manifest.json records row counts and SHA256 values of the Parquet files (`parquet`)
and of the JSONL shards they were converted from (`shards`, `engineering.shards`,
`papers.shards`).

## topic=engineering configuration

`load_dataset(repo, "engineering")` reads `engineering/train-*.parquet`: documents
from general mechanical-engineering venues whose methods transfer to
shipbuilding (fluid dynamics, structures, materials/welding/corrosion,
propulsion/thermal/energy, electrical/control/autonomy, manufacturing and
optimization). Membership is decided by venue alone -- a registered venue
(journal, society, media, technical reference or industry technical pages)
profile or URL pattern, or an arXiv primary category -- as listed in
`engineering_venues_v1` and recorded per row (`venue`, `venue_kind`, `fields`,
`venue_rule`). Rows carry `topic="engineering"`. Only documents the maritime
model scored below the maritime cutoff are included (0.655 until 2026-09-30,
0.4 since); documents already in this configuration stay here and are not
added to `default`, so no document appears in both configurations; `maritime_score` and the run fields are provenance only.

## topic=papers configuration

`load_dataset(repo, "papers")` reads `papers/train-*.parquet`: every scholarly
paper collected by citation expansion from the maritime corpus, kept outside
the main crawl until a final review. No topical filter is applied: expansion
drifts off topic with citation depth (`citation_depth`), and an audit of a
sample found most papers unrelated to maritime engineering, so filter before
use. `maritime_score` is a preliminary signal only; the model was not calibrated
for papers and under-scores related engineering. `review_label` is set when a
validated review labelled the paper `direct_shipbuilding` or
`related_engineering`; `review_state="pending"` means no final review has been
made. Each distinct text appears once (`paper_source="deferred_reference"`,
keyed by `text_sha256`); papers stored as crawled documents
(`paper_source="document"`) appear only when neither `default` nor
`engineering` holds that document. The first 705 rows (policy v1) were the
papers with score >= 0.4 or an accepting review. Rows carry `topic="papers"`,
the resolved work identifiers and the citation depth of the acquisition.

## Parquet files and the tokenized partition

Every configuration is stored as Parquet since 2026-09-30. The JSONL files they
were converted from remain in commit `fb9f75ed5e5e6a679e07783c48ec9f8dfb4ce584` (`revision="fb9f75ed5e5e6a679e07783c48ec9f8dfb4ce584"`).

The `tokenized_<config>` configurations (`tokenized/<config>/train-*.parquet`)
hold, row for row with the text file of the same name, the row id and
`input_ids` (int32) of `text` under the Qwen3.8 tokenizer (`Qwen/Qwen3.8-27B`
revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`, `tokenizer.json` sha256 `0997f410c57a`).
No special tokens are added; separate documents with `<|endoftext|>` (id 248044)
when packing. The tokenizer normalizes to NFC, so decoding returns the NFC form
of `text`. The row id is `document_id`; papers use `document_id` and `job_id`
together, of which exactly one is set.

| config | rows | tokens | files |
|---|---:|---:|---:|
| `default` / `tokenized_default` | 6,759,973 | 6,496,715,754 | 61 |
| `engineering` / `tokenized_engineering` | 251,641 | 777,978,232 | 7 |
| `papers` / `tokenized_papers` | 12,670 | 152,004,308 | 2 |

<!-- release-stats:start -->
## Release v1.0 (2026-09-29)

Tag `v1.0` fixes the data files of commit `996ade125984e7708d48645ce08210ba2ddf6040`; the commit it points to
changes only this card. The stream keeps appending to `main` after the tag, so
load this version with `revision="v1.0"`.

| config | rows | shards |
|---|---:|---:|
| `default` (maritime, score >= 0.655) | 4,547,716 | 768 |
| `engineering` (venue lane, maritime score < 0.655) | 250,544 | 56 |

`default` = the 1,421,861-row original snapshot plus 3,125,855 rows from completed
filtering windows; `engineering` follows policy `venue-registry-below-maritime-cutoff-v1` with
registry `engineering_venues_v1` (sha256 `a0ca32eaa1ca`).
Documents still awaiting a score or an arXiv category at tag time are not included.

Engineering rows by venue kind:

| venue_kind | rows |
|---|---:|
| media | 137,243 |
| society | 44,143 |
| journal | 40,397 |
| preprint | 21,873 |
| industry | 5,001 |
| reference | 1,887 |
Engineering rows by field (a row counts once per field it lists):
| field | rows |
|---|---:|
| manufacturing_optimization | 182,266 |
| electrical_control_autonomy | 146,629 |
| fluid_dynamics | 121,681 |
| materials_joining_corrosion | 98,975 |
| structural_mechanics | 76,978 |
| thermal_energy | 56,200 |
| propulsion | 9,485 |
Largest engineering venues:
| venue | rows |
|---|---:|
| Design World | 53,915 |
| ASME Topics & Resources | 43,480 |
| AZoM news | 34,629 |
| AZoM articles | 15,574 |
| Engineering.com | 12,330 |
| The Robot Report | 9,941 |
| Machine Design | 7,354 |
| arXiv math.OC | 6,777 |
| Frontiers in Energy Research | 5,325 |
| arXiv cs.RO | 4,853 |
| Materials Transactions (JIM) | 4,702 |
| Frontiers in Materials | 3,885 |
| arXiv cond-mat.mtrl-sci | 3,443 |
| COMSOL Blog & Learning Center | 3,401 |
| ISIJ International | 3,288 |
| Korean Journal of Metals and Materials (대한금속·재료학회지) | 3,205 |
| arXiv eess.SY | 3,071 |
| Frontiers in Robotics and AI | 2,688 |
| Journal of Welding and Joining | 2,672 |
| Fluid Power World | 2,167 |
| Frontiers in Built Environment | 2,122 |
| Journal of the Korean Society of Manufacturing Process Engineers | 1,721 |
| arXiv physics.flu-dyn | 1,603 |
| Gear Technology | 1,333 |
| Journal of the Korean Society for Fluid Machinery | 1,231 |
| Mechanical Sciences (Copernicus) | 1,124 |
| Total Materia Articles | 1,040 |
| Tetsu-to-Hagane (鉄と鋼) | 899 |
| Frontiers in Mechanical Engineering | 865 |
| FORCE Technology Insights | 711 |
Source content retains its original rights; no blanket license is assigned.
<!-- release-stats:end -->

	---
	license: other
	license_name: source-rights-retained
	pretty_name: Haebim Maritime Filtered Corpus
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data/train-*.parquet
	- config_name: engineering
	data_files:
	- split: train
	path: engineering/train-*.parquet
	- config_name: papers
	data_files:
	- split: train
	path: papers/train-*.parquet
	- config_name: tokenized_default
	data_files:
	- split: train
	path: tokenized/default/train-*.parquet
	- config_name: tokenized_engineering
	data_files:
	- split: train
	path: tokenized/engineering/train-*.parquet
	- config_name: tokenized_papers
	data_files:
	- split: train
	path: tokenized/papers/train-*.parquet
	---
	# Haebim maritime filtered corpus

	Parquet (zstd), converted on 2026-09-30 from the JSONL shards uploaded directly from
	PostgreSQL using the Hugging Face API; rows, order and values are unchanged.
	Selection: maritime v2 score >= 0.4 (>= 0.655 until 2026-09-30; filter
	`maritime_score >= 0.655` to reproduce it), 1024-token runs recorded in manifest.json.
	Unscored/rejected and legacy 384-token results are excluded. Full decoded stored
	text is included without repair or truncation; scores used at most 4096 characters
	and 1024 tokens. Each row preserves URL, source, crawl time, score time, model,
	bundle digest, run ID, backend and text SHA256. Cross-URL duplicates are retained.
	Scores are machine predictions; the cutoff is not human-gold calibrated.
	Source content retains its original rights; no blanket license is assigned.
	The original snapshot is preserved. Completed filtering windows were appended at score >= 0.4 until collection stopped on 2026-09-30. Document IDs are deduplicated across original and incremental uploads. Reference-only papers awaiting final review are published separately in the `papers` configuration. The train split is a
	corpus partition, not a held-out evaluation set.

	manifest.json records row counts and SHA256 values of the Parquet files (`parquet`)
	and of the JSONL shards they were converted from (`shards`, `engineering.shards`,
	`papers.shards`).

	## topic=engineering configuration

	`load_dataset(repo, "engineering")` reads `engineering/train-*.parquet`: documents
	from general mechanical-engineering venues whose methods transfer to
	shipbuilding (fluid dynamics, structures, materials/welding/corrosion,
	propulsion/thermal/energy, electrical/control/autonomy, manufacturing and
	optimization). Membership is decided by venue alone -- a registered venue
	(journal, society, media, technical reference or industry technical pages)
	profile or URL pattern, or an arXiv primary category -- as listed in
	`engineering_venues_v1` and recorded per row (`venue`, `venue_kind`, `fields`,
	`venue_rule`). Rows carry `topic="engineering"`. Only documents the maritime
	model scored below the maritime cutoff are included (0.655 until 2026-09-30,
	0.4 since); documents already in this configuration stay here and are not
	added to `default`, so no document appears in both configurations; `maritime_score` and the run fields are provenance only.

	## topic=papers configuration

	`load_dataset(repo, "papers")` reads `papers/train-*.parquet`: every scholarly
	paper collected by citation expansion from the maritime corpus, kept outside
	the main crawl until a final review. No topical filter is applied: expansion
	drifts off topic with citation depth (`citation_depth`), and an audit of a
	sample found most papers unrelated to maritime engineering, so filter before
	use. `maritime_score` is a preliminary signal only; the model was not calibrated
	for papers and under-scores related engineering. `review_label` is set when a
	validated review labelled the paper `direct_shipbuilding` or
	`related_engineering`; `review_state="pending"` means no final review has been
	made. Each distinct text appears once (`paper_source="deferred_reference"`,
	keyed by `text_sha256`); papers stored as crawled documents
	(`paper_source="document"`) appear only when neither `default` nor
	`engineering` holds that document. The first 705 rows (policy v1) were the
	papers with score >= 0.4 or an accepting review. Rows carry `topic="papers"`,
	the resolved work identifiers and the citation depth of the acquisition.

	## Parquet files and the tokenized partition

	Every configuration is stored as Parquet since 2026-09-30. The JSONL files they
	were converted from remain in commit `fb9f75ed5e5e6a679e07783c48ec9f8dfb4ce584` (`revision="fb9f75ed5e5e6a679e07783c48ec9f8dfb4ce584"`).

	The `tokenized_<config>` configurations (`tokenized/<config>/train-*.parquet`)
	hold, row for row with the text file of the same name, the row id and
	`input_ids` (int32) of `text` under the Qwen3.8 tokenizer (`Qwen/Qwen3.8-27B`
	revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`, `tokenizer.json` sha256 `0997f410c57a`).
	No special tokens are added; separate documents with `<\|endoftext\|>` (id 248044)
	when packing. The tokenizer normalizes to NFC, so decoding returns the NFC form
	of `text`. The row id is `document_id`; papers use `document_id` and `job_id`
	together, of which exactly one is set.

	\| config \| rows \| tokens \| files \|
	\|---\|---:\|---:\|---:\|
	\| `default` / `tokenized_default` \| 6,759,973 \| 6,496,715,754 \| 61 \|
	\| `engineering` / `tokenized_engineering` \| 251,641 \| 777,978,232 \| 7 \|
	\| `papers` / `tokenized_papers` \| 12,670 \| 152,004,308 \| 2 \|

	<!-- release-stats:start -->
	## Release v1.0 (2026-09-29)

	Tag `v1.0` fixes the data files of commit `996ade125984e7708d48645ce08210ba2ddf6040`; the commit it points to
	changes only this card. The stream keeps appending to `main` after the tag, so
	load this version with `revision="v1.0"`.

	\| config \| rows \| shards \|
	\|---\|---:\|---:\|
	\| `default` (maritime, score >= 0.655) \| 4,547,716 \| 768 \|
	\| `engineering` (venue lane, maritime score < 0.655) \| 250,544 \| 56 \|

	`default` = the 1,421,861-row original snapshot plus 3,125,855 rows from completed
	filtering windows; `engineering` follows policy `venue-registry-below-maritime-cutoff-v1` with
	registry `engineering_venues_v1` (sha256 `a0ca32eaa1ca`).
	Documents still awaiting a score or an arXiv category at tag time are not included.

	Engineering rows by venue kind:

	\| venue_kind \| rows \|
	\|---\|---:\|
	\| media \| 137,243 \|
	\| society \| 44,143 \|
	\| journal \| 40,397 \|
	\| preprint \| 21,873 \|
	\| industry \| 5,001 \|
	\| reference \| 1,887 \|

	Engineering rows by field (a row counts once per field it lists):

	\| field \| rows \|
	\|---\|---:\|
	\| manufacturing_optimization \| 182,266 \|
	\| electrical_control_autonomy \| 146,629 \|
	\| fluid_dynamics \| 121,681 \|
	\| materials_joining_corrosion \| 98,975 \|
	\| structural_mechanics \| 76,978 \|
	\| thermal_energy \| 56,200 \|
	\| propulsion \| 9,485 \|

	Largest engineering venues:

	\| venue \| rows \|
	\|---\|---:\|
	\| Design World \| 53,915 \|
	\| ASME Topics & Resources \| 43,480 \|
	\| AZoM news \| 34,629 \|
	\| AZoM articles \| 15,574 \|
	\| Engineering.com \| 12,330 \|
	\| The Robot Report \| 9,941 \|
	\| Machine Design \| 7,354 \|
	\| arXiv math.OC \| 6,777 \|
	\| Frontiers in Energy Research \| 5,325 \|
	\| arXiv cs.RO \| 4,853 \|
	\| Materials Transactions (JIM) \| 4,702 \|
	\| Frontiers in Materials \| 3,885 \|
	\| arXiv cond-mat.mtrl-sci \| 3,443 \|
	\| COMSOL Blog & Learning Center \| 3,401 \|
	\| ISIJ International \| 3,288 \|
	\| Korean Journal of Metals and Materials (대한금속·재료학회지) \| 3,205 \|
	\| arXiv eess.SY \| 3,071 \|
	\| Frontiers in Robotics and AI \| 2,688 \|
	\| Journal of Welding and Joining \| 2,672 \|
	\| Fluid Power World \| 2,167 \|
	\| Frontiers in Built Environment \| 2,122 \|
	\| Journal of the Korean Society of Manufacturing Process Engineers \| 1,721 \|
	\| arXiv physics.flu-dyn \| 1,603 \|
	\| Gear Technology \| 1,333 \|
	\| Journal of the Korean Society for Fluid Machinery \| 1,231 \|
	\| Mechanical Sciences (Copernicus) \| 1,124 \|
	\| Total Materia Articles \| 1,040 \|
	\| Tetsu-to-Hagane (鉄と鋼) \| 899 \|
	\| Frontiers in Mechanical Engineering \| 865 \|
	\| FORCE Technology Insights \| 711 \|

	Source content retains its original rights; no blanket license is assigned.
	<!-- release-stats:end -->