CoolFace
Datasetpublic

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… See the full description on the dataset page: https://huggingface.co/datasets/anuj-inavlabs/Thinkspark-v2-270m-training-data.

sourceHugging Faceapache-2.0updated 22d agoView on Hugging Face
0likes6.9kdownloads
Dataset Card

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:

columntypedescription
audioAudioplayable wav (already stored under audio/)
user_textstringthe exact line synthesised (Section 8.4)
scenario_idstringstable id, joins to timestamps/<shard>/<id>.json
behaviourstringone of the 12 generation buckets (Section 8.1)
languagestringhi / en / gu / hiennative / guennative
domainstringbfsi_collections / support / sales
genderstringrequested TTS voice gender
prosodystringfalling / rising / held / flat / distressed / neutral
agent_textstringwhat the agent was saying (may be empty)
duration_sfloat64audio duration in seconds
num_wordsint64word count in the Soniox timestamp alignment

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

python
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).

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