rushilrawat/garhwali-corpus

Garhwali Language Lab

Current developer status — 8 October 2026

The live Dataset Viewer reports 20 configurations, 31 config/split views, and 963,484 displayed rows. These views overlap and include source indexes; they are not 963,484 distinct training examples. The screened_meta_gbm view has 1,841 automatically screened sentence-length transcripts for exploratory text training, and short_utterances_meta_gbm has 110 context rows. Both reuse transcript values already present in older configurations. No new source text or recording was added by these views, and no native-speaker review has been completed.

For a Python 3.12 setup and CPU-first masked-language-model example, use the GitHub developer quick start. The readiness guide explains the training and evaluation limits. The current platform audit separates source records, content views, and training candidates. GitHub CI passed 832 tests on 8 October; this is a tooling check, not a model-quality score.

Release: garhwali-language-lab-v0.2.8

Size and training-readiness snapshot — 6 October 2026

The live Hub repository displays 963,484 rows and 7.57 GB of files. This is not a count of unique language examples: 778,157 rows (80.8%) are source/reference tables, and the remaining 185,327 are overlapping config views. The published text config has 18,598 rows / 291,914 whitespace- separated words / 1,373,045 characters. Across the historical v0.2.7 text, text_expansion (1,737 rows), and text_resources (475 rows) configs, zero rows are marked recommended for general text-model training. The separate v0.2.8 screened_meta_gbm config adds 1,841 experimental training candidates; see its quality fields and review limitations below. The 14,988 paharili_gbm records are experimental language-identification material with unresolved sentence origins and unreviewed labels. The 2,002-row ASR view is speaker-disjoint but unreviewed. The linked speech repository is separate (113,363 rows / 36.5 GB); SraVaani outputs are machine drafts, not gold transcripts. See the metrics and utility audit and the project source copy.

Clean screened text config

The v0.2.8 addition exposes two purpose-labeled configs. screened_meta_gbm contains 1,841 non-empty, normalized-deduplicated Garhwali transcript texts suited to sentence-level experimental LM training (47,566 whitespace-separated words). short_utterances_meta_gbm preserves 110 short, nonblank utterances for context; they are not recommended for general LM training. All 1,951 selected text values were found verbatim in the earlier v0.2.7 text/text_expansion configs. This release adds a source-decided, quality-filtered training surface, not new unique transcript content. Blank values, normalized duplicates, reference tables, machine ASR drafts, and very short utterances are not in the training config. Each row keeps source record IDs, attribution, rights evidence, and quality fields. This is automated screening, not native-speaker review or verified linguistic ground truth.

Configuration	Rows	Use
screened_meta_gbm	1,841	Experimental sentence-level text training candidates
short_utterances_meta_gbm	110	Short-utterance context; not general LM training

from datasets import load_dataset
data = load_dataset("rushilrawat/garhwali-corpus", "screened_meta_gbm", split="train")
print(data[0]["text"])

Developer quick start

Install the small tabular-data stack with pip install datasets pandas duckdb, then stream a tiny vocabulary sample:

from datasets import load_dataset

lexicon = load_dataset(
    "rushilrawat/garhwali-corpus", "lexicon", split="train", streaming=True
)
for row in lexicon.take(3):
    print(row["form"], row.get("glosses"), row["rights_status"], row["quality_status"])

The common record envelope is rights_status, reuse_scope, license_labels, quality_status, and record_quality_flags; source-specific provenance and review fields remain alongside it. Dataset cards and schema guide explain how to filter these fields and how to read each configuration.

Configurations and current row counts

Counts are rows in this release's views, not unique examples. Configurations can overlap, and reference-index rows are not training examples.

Configuration	Rows	Use
asr	2,002	provider/human transcripts (unadjudicated)
catalog	36,105	unique text inventory with content and metadata-only entries
geography	50	place facts and citations
historical_terms	36	historical names and terms
instructions	3,228	instruction/response examples
lexicon	1,493	vocabulary and pronunciation candidates
literary_people	26	writer and contributor metadata
literary_works	66	work-level bibliography
paharili_gbm	14,988	PahariLI Garhwali-labeled text-classification examples; sentence origins and rights remain unresolved
popular_songs	30	song-level metadata, no lyrics
sravaani_drafts	104,534	machine transcript drafts; not ground truth
text	18,598	Garhwali text examples
text_expansion	1,737	strict-tier train-split candidates; zero currently meet training recommendation; not evaluation data
text_resources	475	rights-cleared supplementary text; varied quality, not for evaluation
university_research	8	research bibliography
record_index	355,846	one metadata reference per archived row; not training data
source_catalog	9,571	deduplicated source and rights references
record_sources	412,740	record-to-source join table
screened_meta_gbm	1,841	experimental sentence-level text training candidates
short_utterances_meta_gbm	110	short-utterance context; not general LM training

Pandas and DuckDB recipes, the full schema, searchable lexicon example, and speech loading instructions are in the developer quick start and schema guide. A minimal lexicon search script is shipped at the release root.

What this repository provides

The content views contain 185,327 overlapping rows across 17 named content configurations and 28 content config/split views. With three reference configurations, the full dataset has 20 named configurations and 31 config/split views. Reuse terms vary by row and configuration. The paharili_gbm config is an explicitly labeled experimental inclusion: PahariLI declares an Apache-2.0 repository license, while sentence-level source origins remain unresolved. Separately, the metadata-only reference index covers every one of the 355,846 rows in the complete all-data archive, including all 216 structured geography, history, literature, music, and research records.

The reference index includes 9,571 deduplicated source records and 412,740 record-to-source links. It exposes available names and titles, source URLs, attribution, rights status, and quality metadata. Record-level rights are marked not_recorded for 0 rows; other rows include rights-pending, reviewed-unresolved, metadata-only, or cleared statuses. not_recorded is not permission to reuse a record: inspect its linked source entry and terms. The index does not contain the referenced works' text, lyrics, transcripts, audio, speaker identifiers, local paths, or content hashes. These are archive-row counts with overlapping views, not unique-example counts; the reference rows themselves are not training examples. Links and bibliographic facts do not grant rights to copy or reuse source material.

All 216 structured geography, history, literature, song, and research records appear in factual/bibliographic form; no records are dropped from these metadata configurations. Prose notes, lyrics, translations, abstracts, and source passages are omitted unless separately licensed. For retained catalog entries, the full text not included in this public profile remains in the local all-data package; entries whose source text was not acquired remain represented by their available source metadata. Each text-catalog record carries its specific rights state: CC BY-SA rows require attribution and share-alike; CC BY-NC-SA rows are noncommercial and share-alike; the five PIB instrument terms cite the PIB reproduction policy; and the 193 Mountain Voices glossary headwords carry Panos's attributed-reproduction guideline for press, educational/research institutions, and nonprofits. That guideline does not expressly address commercial scope or model training, so those values are excluded from model-training views. The 16 isolated one-token facts are listed without definitions, source record positions, or list ordering. Lexical facts from unlicensed thematic sources are included only when independently present in at least two distinct source collections; all such facts remain catalog-only, outside training views, and retain language-review flags. See the source-by-source rights-resolution log. Native-speaker review and dialect annotation are deferred; benchmark and model scores are automated research results, not native-validated claims.

Project history

This release is the result of a staged corpus build, with older releases kept available under their own versioned paths:

1. v0.1.x — establish the corpus workflow. The project began by collecting Garhwali text and language references with source attribution, then added a reproducible preparation and release pipeline.
2. v0.2.0 — expand Garhwali-only sources. A deduplicated source wave added 671 exact-unique texts, including historical Garhwali specimens and stories. Historical OCR and translated stories remain visibly marked as unreviewed.
3. v0.2.1 — continue the corpus and make it easier to use (30 September–1 October 2026). The follow-up added 164 exact-new strings after deduplication, brought more values into the rights-filtered profile under their recorded bases, and exposed factual/bibliographic metadata for all 216 structured records. The release also adds a consistent rights-and-quality envelope, exact-count quick start, schema guide, searchable lexicon example, and linked speech dataset. Earlier release files remain available.

The figures below describe the current v0.2.8 package, not a cumulative sum of overlapping views. Row-level terms and quality labels remain authoritative.

This public profile contains Garhwali text, human transcripts, lexicon and instructions, plus experimental SraVaani drafts. The geography, history, literature, song, and university-research configurations contain factual and bibliographic metadata, with expressive notes, lyrics, summaries, and source passages omitted.

This public-profile package contains 185,327 overlapping content-view rows across 17 content configs. The three reference-table configs add 778,157 overlapping-view rows, for 20 named configs and 31 config/split views total. These counts are views, not unique examples. The SraVaani draft config covers 104,534 unique audio hashes from 104,542 source rows: 104,500 have non-empty draft text and 34 are empty. This transcript-only package does not include audio files or source filenames.

This dataset card describes data scope, not model accuracy. Native-speaker review and dialect annotation are deferred; transcripts and machine drafts retain their review status. The project pipeline uses Python 3.12 and LangGraph for resumable ingestion; quality checks and model metrics use task-specific scripts. LangChain is not part of the current pipeline. See the project README and benchmark research status for the latest measured results and limitations.

The catalog configuration publicly accounts for all 36,105 exact-unique collected text records. Rows whose source terms do not permit redistribution retain their stable content hash, source URL, rights status, quality tier, language evidence, and review reasons; their full text is not included in public content. The current catalog exposes 134 CC BY-NC-SA 4.0 records, 530 total share-alike records, and 198 entries under source-specific reproduction policies (5 PIB instrument facts and 193 Mountain Voices glossary headwords), plus 16 exact one-token facts published without definitions, source record positions, or list arrangement. The fact-only projection includes individually selected tokens and lexical tokens that appear in at least two distinct thematic source collections; it is not a copy of any source list. The Panos guideline permits attributed reproduction by press, educational/research institutions, and nonprofits; commercial scope and machine-learning training are not expressly addressed, so these headwords are excluded from model-training views. Per-row terms apply; the package asserts no blanket content license.

Upstream source-split overlap

The text package preserves every row. The text/source_overlap split contains 1,305 rows previously assigned to text/train that carry a Meta dev/test record ID, an official VAANI test-remainder source record, or a transcript match to held-out Meta/VAANI source material. The text_expansion/source_overlap split contains 454 similar catalog additions. These rows remain downloadable and retain their source metadata, but are kept outside the default train split. This is a conservative source-lineage warning, not a claim that all matched audio recordings are identical. See the split-overlap audit.

Fast-tracked text expansion

The text_expansion config adds 1,737 Garhwali candidate texts from the existing catalog. Each passed automated strict-tier and recorded rights-basis checks, is absent from the existing text splits after NFKC/alphanumeric normalization, and has no source-record ID matching existing text validation/test or frozen benchmark validation/test rows. 454 values with upstream Meta or VAANI development/test source matches remain published in the source_overlap split, outside default train. The values remain machine-screened, not native-reviewed. 0 currently meet the project's conservative training-recommendation rule; 1,737 do not pass the current source-eligibility or quality gates. This is not training approval; source-page-family isolation is not assessed, so it is not an independent evaluation set. Values also appear in the catalog inventory by design; do not add config row counts when reporting unique texts.

Supplementary text resources

The text_resources config adds 475 additional normalized-unique Garhwali records (43,518 whitespace-delimited words; 196,400 characters) already present in the source catalog. These pass the recorded redistribution-basis and Garhwali-language filters and are absent from the existing text-bearing views after normalized deduplication. Their quality tiers vary; 0 are currently recommended for training and none are recommended for evaluation. The config is intended for lookup and research. Check each row's redistribution_status, license, quality flags, and source terms before reuse. CC BY-NC-SA rows are limited to noncommercial use and require share-alike; this subset is not uniformly training-ready or a native-reviewed text set.

PahariLI Garhwali text addition

The new paharili_gbm config adds 14,988 normalized-unique Garhwali-labeled sentence records from the PahariLI corpus. It starts from 15,000 upstream GBM-labeled rows, which reduce to 14,989 distinct normalized texts after collapsing 11 repeated source rows. All source record IDs and surface variants are retained: 14,999 source IDs appear in this config, and 1 remain on the exact-matching row already in text_expansion. That existing text is not duplicated here.

The config keeps the upstream language-identification splits: 11,989 train, 2,995 test, and 4 text groups that occur in both upstream splits. The last group stays in source_overlap to avoid train/test leakage. This test split is the upstream PahariLI classification split; it is not an independent evaluation set for the project's language model or other tasks.

The PahariLI repository includes an Apache-2.0 license file, but its README does not identify the sentence-level source of each item. Each row therefore keeps the repository-level license declaration, raw-file URL and SHA-256, source index, attribution, and the existing source_lineage_missing, component_rights_review_required, and possible scripture/blog flags. The release makes no claim that Apache-2.0 independently clears underlying source text. Rows are unreviewed and not recommended for general language-model training; the upstream train partition remains usable for experimental language-identification work. The upstream license text is included at licenses/PahariLI-Apache-2.0.txt.

The asr configuration contains human transcripts from VAANI. The sravaani_drafts configuration contains machine-generated hypotheses from ARTPARK-IISc/SraVaani-1.0 revision f5dd5358325a5208775b91dad98918e079ea2b27; these are noisy experimental data, not human ground truth. Targeted rows also retain their local Whisper alternative, cross-model agreement, and bounded review-confidence evidence. Draft export status: complete.

All 1,188 targeted recordings retain the third-checkpoint hypothesis. For the 1,090 Garhwali review records, the package also carries a structurally ranked machine proposal and its evidence limits. These proposals are pending audio review and are never represented as automatic corrections or human references. The remaining 35 structural outliers also include deterministic re-decoding, waveform activity, and human-reference calibration evidence. They still require listening review.

The draft layer also preserves 98 source-label conflict recordings. These remain available for auditing but are explicitly ineligible for Garhwali training.

To keep Dataset Viewer schemas stable across shards, the SraVaani draft configuration stores nested evidence objects and draft quality-flag arrays as compact JSON strings. Parse those columns with a JSON parser; their values are preserved without flattening or omission.

This package uses multiple upstream licenses. Inspect each row's provenance before redistribution or model release. Full documentation, limitations, and the release audit are in the source repository.

Complete source reference index

The three metadata tables cover 355,846 records from the complete local all-data archive. record_index has one row per archived record, source_catalog contains 9,571 deduplicated source references, and record_sources contains 412,740 join rows. The tables include source URLs, available attribution, rights and quality status, and factual catalog fields. They exclude source text, lyrics, transcripts, machine drafts, audio, speaker identifiers, local paths, and content hashes.

The v0.2.7 baseline had 183,376 rows across its content configurations; v0.2.8 adds 1,951 quality-screened text rows, bringing the displayed content configuration total to 185,327; 174,820 indexed records have a corresponding public content value. Some catalog entries retain identifiers and source/rights metadata while their full text remains outside the catalog config; PahariLI text is separately exposed in paharili_gbm with its unresolved origin status. Reference-index rows are not training examples and must not be added to corpus sample counts. Source links point to original material and do not grant reuse rights.

Text availability and source lookup

The catalog config has 23,448 blank text fields. A cross-config scan of all public JSONL string values found 15,004 of those normalized texts elsewhere in this repository (14,988 in paharili_gbm and 16 in other fields). 8,444 distinct texts are not represented in any public config; the project retains them in its local all-data package with rights_pending / not_cleared status. Each has a source locator: 7,675 rows contain an inline URL and 769 resolve a URL through the public source_catalog. Some locators identify a collection or dataset rather than the exact work or page. The public developer quick start shows the lookup. A source link is a locator, not a reuse license; a matching PahariLI string is not rights clearance. See the full availability audit.

See reference_index_manifest.json for counts by family and the three table checksums.

Linked speech dataset

The companion Garhwali Speech dataset contains VAANI audio with provider transcripts and separately labeled SraVaani drafts.