yashwardhan20417/codemixed-kashmiri

Codemixing2 generated transliteration dataset

This folder combines the generated outputs from the six Codemixing2 runs into one row-aligned dataset. It contains 24,402 unique Kashmiri transcript rows, ordered by row_id from 0 through 24,401.

Files

• dataset.parquet — recommended dataset file; includes the source transcript, word-level dictionary sequence, four generated sentence forms, per-row generation provenance, and audit assessments.
• dataset.jsonl — the same rows in JSON Lines format. word_sequence and audit are nested JSON values here.
• merge_manifest.json — run coverage, prompt hashes, and combined audit verdict counts.

Fields

• row_id — position in the exact-text-deduplicated IndicVoices-R Kashmiri training transcript set used for generation. This is not a raw IndicVoices-R row number.
• text — original Kashmiri transcription provided to generation.
• word_sequence / word_sequence_json — ordered logical-word entries and their word-level mappings. It is a JSON value in dataset.jsonl and a JSON-encoded string in Parquet.
• word_count, source_token_count — counts from the generation output.
• partial_transliteration, full_transliteration, full_transliteration_with_diacritics, full_transliteration_with_correct_stem — the four generated sentence forms.
• generation_* — provider, model, reasoning effort, source run, and generation-prompt SHA-256 for that row. The run used Claude Sonnet 5.5 High for most rows through 12,999 and GPT-6 Luna High for later rows; the later Luna run also filled 60 row IDs missing from earlier run outputs.
• audit_* — GPT-6 Sol Medium model assessments for the four generated forms. In Parquet, each form has a verdict, confidence, and JSON-encoded issues list, alongside audit_json. In JSONL, assessments are nested under audit.

Scope and interpretation

This is the generated transcript dataset only. It does not include audio or the original IndicVoices-R metadata, and it has not been merged back into all 25,095 raw IndicVoices-R rows. Duplicate source transcripts were collapsed before generation, so the dataset has 24,402 rows.

Audit verdicts are model assessments, not human-verified gold labels. Correct-stem corrections were applied to generated sentence outputs in the source run artifacts. The word_sequence field preserves the word-level mappings and was not regenerated from those sentence edits, so a manually corrected correct-stem sentence may not exactly reconstruct from its stored word mappings. Some originally flagged correct-stem rows were re-audited and those verdicts were merged into the run audits; this was not a full re-audit of every corrected row. The other audit fields retain their stored audit verdicts.

Source and attribution

The source transcripts are from IndicVoices-R, whose dataset card lists the license as CC-BY-4.0. This derived text dataset is released under CC-BY-4.0 with attribution to AI4Bharat and the IndicVoices-R authors. Please cite the source work:

Ashwin Sankar et al. “IndicVoices-R: Unlocking a Massive Multilingual Multi-speaker Speech Corpus for Scaling Indian TTS.” NeurIPS 2024 Datasets and Benchmarks Track.