AumPandya01/unspoken-data

Unspoken hackathon data

Data created by team Unspoken at the NVIDIA + Red Hat Small Models Hack (Raleigh, 17–18 Oct 2026): synthetic and derived training/evaluation sets, plus recordings of the team members themselves. Code and full methodology: https://github.com/Aum3010/unspoken-nvidia-small-models-hack (docs/DATA_RELEASE.md, results in docs/RESULTS.md).

Contents

Folder	Rows	What	How it was made
coordinator_gym/	12 example, 72 eval, 44 heldout	NeMo Gym unspoken_communication tasks: agent candidates + quality flags → one coordinator tool call	Fully synthetic. evaluations/nemo_gym/unspoken_communication/create_examples.py, seeds 1 / 7 / 2026; heldout drops every task also in eval
neuro_comm_gym/	5 example, 600 train, 120 eval	NeMo Gym neuro_comm tasks: when to stop flashing and commit a letter	evaluations/nemo_gym/neuro_comm/create_examples.py --seed 7, from bigP3BCI Study F evidence
neuro_comm_student/	1,717 hindsight, 8,394 value	LoRA labels for Nemotron Nano 4B on neuro_comm (state text → more / commit / abstain)	scripts/benchmark/neuro_comm_student.py data (hindsight, attempt 1) and value_data (expected Gym reward, attempt 2); bigP3BCI Study F calibration sessions 1–2 only
coordinator_student/	700 per file	E7b/E7c coordinator router: voice + lips phrase scores and EEG option → action	scripts/benchmark/coordinator_student.py prep; GRID s1/s2 scores from our E1 run, 2-fold by speaker
lip_enrollment/	22 phrases, 90 recordings	Auto-AVSR encoder features (NNN.npy, frames × 768, float16) + phrase.txt; no video	One team member mouthing the patient phrases, recorded in the app's enrollment flow
voice_recordings/	307 WAV, 3 speakers	16 kHz mono; {speaker}_{condition}_{phraseid}_{take}.wav; ids and conditions in phrases_and_conditions.yaml	Team members p01 (normal, noisy, whisper, muffled), p02 (normal, whisper), p03 (normal; many clips clipped, recorded too loud). 20 phrases, 7 hard negatives, silence / breathing / cough / unrelated

Consent

Every person recorded is a member of the team; all agreed on 18 Oct 2026 to publish these recordings openly under CC BY 4.0. Speakers are pseudonymous (p01–p03). Do not attempt to identify them.

Sources and attribution

• bigP3BCI v1.0.0, PhysioNet, https://doi.org/10.13026/0byy-ry86, CC BY 4.0 (Study F, people with ALS). neuro_comm_gym and neuro_comm_student are derived from it.
• GRID audiovisual corpus, Cooke, Barker, Cunningham & Shao (2006), CC BY 4.0. coordinator_student is derived from model scores on speakers s1 and s2.
• Models used to produce the data: NVIDIA Nemotron Speech Streaming en 0.6B (voice scores), Auto-AVSR (lip features and scores).

Limits

• Small: three speakers, one lip-enrollment speaker. None of the recorded people has a speech impairment.
• The Gym tasks are synthetic and the student labels are model-made; they are training/evaluation material for this prototype, not clinical data.