README.md

---
pretty_name: Robin186 Unified Indian Sign Language Vocabulary
license: other
language:
- en
tags:
- indian-sign-language
- accessibility
- vocabulary
configs:
- config_name: default
  data_files:
  - split: train
    path: data/isl.jsonl
---

# Robin186/ISL

Public, provenance-first vocabulary index for Indian Sign Language. The 2026-07-17
snapshot combines the public ISLRTC A-Z dictionary index with a small local INCLUDE
metadata snapshot, a validated motion catalog, and narrowly filtered SignSaarthi
lexicon annotations. It does not
download or contain dictionary videos, raw video corpora, audio, or signer media.

## Coverage

| Measure | Count |
| --- | ---: |
| Dataset records | 12434 |
| Official ISLRTC entries preserved | 15560 |
| Canonical official terms | 12357 |
| Official duplicate-term groups consolidated | 2979 |
| INCLUDE vocabulary labels | 263 |
| Records with INCLUDE motion metadata | 262 |
| INCLUDE labels quarantined from motion playback | 1 |
| Official terms with INCLUDE motion metadata | 186 |
| Locally enriched admitted records | 12434 |

Repeated official terms are one dataset row with every distinct official entry ID and
detail URL retained under `official.entries`; they are not silently discarded.

## Record fields

- `id`, `term`, `normalizedTerm`, `language`: stable vocabulary identity.
- `datasetIds`, `sourceUrls`, `official`, `include`: primary provenance, official
  variants, source-split presence, and honest motion availability.
- `motion`: compact INCLUDE catalog metadata. The source catalog contains no landmark
  frame arrays, so `framePayloadIncluded` is always `false` in this bundle.
- `lexicon`: safe local annotations only after the term independently matches ISLRTC
  or INCLUDE. Aliases, token weights, source fields, and timestamps are not copied.
- `license`, `review`: per-row rights and expert-review status without certification claims.

## Licensing and review

This dataset card uses `license: other` because it combines source-specific terms.
The official ISLRTC FAQ permits dictionary use for research, teaching, and technology
only when the data are not resold or used for profiteering and ISLRTC is properly
acknowledged. INCLUDE is tracked as CC-BY-4.0 based on its source metadata. This public
bundle contains metadata and source links only; it does not grant commercial rights or
permission to redistribute source media or derived landmark payloads.

The upstream INCLUDE metadata contains exact media-path overlap between its published
splits, so those split names are retained for provenance only and are not represented
as a leakage-safe evaluation benchmark. A label with no source-consistent training
path is kept as vocabulary but quarantined from motion playback.

An official dictionary listing establishes vocabulary provenance; it does not prove
that this merge or an INCLUDE motion clip received independent ISL expert review.
Motion review status is copied literally from the catalog and is currently represented
per record.

## iSign exclusion

No iSign-only record or gated CSV content is included. The blended local lexicon may
enrich a term only after that term is independently admitted by ISLRTC or an INCLUDE
motion. Unmatched local terms are counted only in `metadata.json`; their text is not
written. The validator rejects iSign dataset provenance and gated local source paths.

## Files

- `data/isl.jsonl`: deterministic records sorted by normalized term.
- `metadata.json`: input/page hashes, coverage, exclusions, licenses, and validation.
- `README.md`: this Hugging Face dataset card.

The bundle is intended for vocabulary lookup, coverage analysis, and motion
catalog joins. It is not a complete motion corpus, translation benchmark, or substitute
for review by qualified ISL users and experts.