Krish187653/rasa-kashmiri-codemixing-transliterations

RASA Kashmiri code-mixing transliterations

This dataset adds ordered word-level language and script annotations and four reconstructed sentence forms to the Kashmiri transcripts in the AI4Bharat RASA dataset. The original train/test split and every source row are retained.

Configurations

• default contains all 20,443 training rows and all 2,269 test rows. Rows without usable generated output retain their source text and have null annotation fields with a status explaining why.
• audited_complete contains rows with reconstructed forms and a completed GPT-6 Sol medium audit: 20,249 train rows and 2,222 test rows. “Completed” means the audit ran; it does not mean every form was judged correct.

The test generation retry for 47 rows was canceled before it completed. Those rows remain in default with their original text and missing_or_invalid / not_audited statuses.

Fields

Field	Description
row_id	Sequential source row identifier; test IDs continue after the training IDs.
text	Original Kashmiri RASA transcript.
word_sequence	Ordered list of logical dictionary entries. Each entry has source_span, origin, ascii, phonetic, language, script, and converted_source_word. A source span may cover multiple source tokens.
word_count	Number of entries in word_sequence; null if output is unavailable.
source_token_count	Whitespace token count after stripping outer punctuation, following the pipeline tokenizer.
partial_transliteration	Sentence with converted spans rendered in ASCII and native spans retained in Kashmiri script.
full_transliteration	Full ASCII transliteration.
full_transliteration_with_diacritics	Full Latin transliteration with phonetic diacritics.
full_transliteration_with_correct_stem	Full Latin transliteration with the pipeline's correct-stem form.
generation_status	validated when generation passed pipeline validation; otherwise missing_or_invalid.
audit_status	completed, not_audited, or not_reconstructable. This records processing state, not an accuracy label.
manual_correction_count	Number of audit-approved manual sentence corrections applied (training rows only).
dictionary_annotation_count	Number of dictionary-only annotations applied (training rows only).

Unavailable word-level and sentence-form annotations are represented as JSON null. The four transliteration columns are generated annotations, not independent human reference translations. Training annotations include the audit-approved corrections materialized in the local corrected training dataset. Test annotations are generated and audited but were not manually corrected.

Pipeline and review limits

The word-level dictionaries were generated with GPT-6 Luna high. Sentence reconstructions were reviewed with GPT-6 Sol medium. Model audits are automated assessments and are not human-verified ground truth. The audited_complete configuration includes all rows with a completed audit regardless of the audit's verdict.

The source transcripts come from AI4Bharat RASA, which is distributed under CC BY 4.0. The source dataset citation is:

Praveen Srinivasa Varadhan, Ashwin Sankar, Giri Raju, and Mitesh M. Khapra. “Rasa: Building Expressive Speech Synthesis Systems for Indian Languages in Low-resource Settings.” INTERSPEECH 2024.

This text-only export contains no audio or audio metadata. See compile_report.json for exact counts, source hashes, and incomplete row IDs.