anuj-inavlabs/Thinkspark-v2-270m-training-data

ThinkSpark-v2-350M — training data

Full-duplex floor-controller (Section 8) training corpus: playable audio + text, paired for the Dataset Viewer, plus every scenario field (behaviour, language, domain, gender, prosody, agent text) and Soniox character-level timestamps.

Dataset Viewer

Default split is parquet with a real Audio feature — a player renders inline next to the text in the Hub UI:

column	type	description
audio	Audio	playable wav (already stored under audio/)
user_text	string	the exact line synthesised (Section 8.4)
scenario_id	string	stable id, joins to timestamps/<shard>/<id>.json
behaviour	string	one of the 12 generation buckets (Section 8.1)
language	string	hi / en / gu / hi_en_native / gu_en_native
domain	string	bfsi_collections / support / sales
gender	string	requested TTS voice gender
prosody	string	falling / rising / held / flat / distressed / neutral
agent_text	string	what the agent was saying (may be empty)
duration_s	float64	audio duration in seconds
num_words	int64	word count in the Soniox timestamp alignment

Paired rows: 18,894 (only scenarios with both real audio and non-empty text).

from datasets import load_dataset
ds = load_dataset("anuj-inavlabs/Thinkspark-v2-270m-training-data", split="train")
print(ds[0]["user_text"], ds[0]["behaviour"])
# ds[0]["audio"] -> array / sampling_rate / path

Layout

• data/train-*.parquet — Viewer source (audio + text + full scenario metadata)
• audio/<shard>/<scenario_id>.wav — rendered user audio (sharded, ≤1000 files/dir)
• timestamps/<shard>/<scenario_id>.json — Soniox character-level timestamps. Training use only (Section 8.4 frame calibration) — never needed at inference, where a live agent-state flag replaces timing entirely (Section 4.3).
• metadata.jsonl / metadata.csv — AudioFolder-style side index

See ThinkSpark-v2-350M for the full pipeline this data feeds (Phase 1 modality alignment + Phase 2 referee fine-tune on Gemma-3-270M).