Datasetpublic

anuj-inavlabs/kupe-thinkspark

ThinkSpark training corpus Synthetic multilingual training data for ThinkSpark — a tiny model that predicts the human thinking sound / backchannel a voice assistant should emit between STT and the main LLM reply. Generated with Sarvam sarvam-105b (reasoning_effort=low) via the kupe-thinkspark pipeline. Files path rows description corpus/thinkspark_corpus_clean.jsonl 38,338 validated + deduped full corpus corpus/thinkspark_corpus.jsonl 40,090 raw… See the full description on the dataset page: https://huggingface.co/datasets/anuj-inavlabs/kupe-thinkspark.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes38downloads

Dataset Card

ThinkSpark training corpus

Synthetic multilingual training data for ThinkSpark — a tiny model that predicts the human thinking sound / backchannel a voice assistant should emit between STT and the main LLM reply.

Generated with Sarvam sarvam-105b (reasoning_effort=low) via the kupe-thinkspark pipeline.

Files

path	rows	description
`corpus/thinkspark_corpus_clean.jsonl`	38,338	validated + deduped full corpus
`corpus/thinkspark_corpus.jsonl`	40,090	raw generated corpus (includes rejects filtered at build)
`splits/train.jsonl`	30,652	stratified train split
`splits/val.jsonl`	3,843	validation split
`splits/test.jsonl`	3,843	test split
`vocab/label_maps.json`	—	language / intent / register / emotion label maps
`vocab/filler_dictionary.json`	—	per-language filler surface forms

Total clean rows: 38,338

Row schema

Each JSONL line:

json

{
  "input": "user's current utterance",
  "context": "prior conversation transcript (may be empty)",
  "context_langs": ["hi"],
  "language": "hi",
  "script": "Deva",
  "register": "casual",
  "intent": "thinking",
  "emotion": "neutral",
  "filler_type": "sound",
  "filler_candidates": ["हम्म...", "अच्छा"],
  "filler_weights": [0.6, 0.4],
  "notes": "..."
}

Source

Built locally with kupe-thinkspark — mirrors the kupe-tts data generation workflow.

Uploaded: 2026-08-24 06:59 UTC