Conduit claims to have generated the world's largest neuro-language dataset, achieving approximately 10,000 hours of multimodal neural data coupled with aligned speech/text, for the purpose of training "thought-to-text" decoders capable of zero-shot semantic prediction. The primary focus of this report is on the operational and logistical strategies required to achieve this scale in a high-throughput, 20-hour-per-day basement facility. Key lessons include aggressively optimizing participant turnover time (down to 3 minutes), implementing dynamic booking systems to maximize booth occupancy, and leveraging data quantity to overcome traditional data quality constraints. As far as I can tell, the central, and efficient, finding is that above 4,000–5,000 hours, sophisticated noise reduction efforts provide marginal utility, allowing the operation to shed expensive quality rituals (like gel application) in favor of maximum volume.
Collecting 10,000 Hours of Neuro-Language Data: Operational and Scalability Lessons
0:00 Ambitious Goal: The project aims to train thought-to-text models that decode semantic content from noninvasive neural data before the subject speaks or types. Zero-shot generalization to new subjects is claimed, which is a surprising result given typical dataset limitations.
Participant Engagement (Content Quality): Participants engage in two-hour freeform conversations with an LLM (typing or speaking).
Initial participant outputs were often incoherent (e.g., repeating keyboard strikes). Quality was addressed by introducing a token quantity/quality scoring system and personalizing LLM interaction.
Operational fixes included physically "simplifying" keyboards to prevent crashes and addressing participant discomfort (ventilation via 8m ceiling pipes; custom padding for 4-pound helmets).
Completion rate drastically improved from under 20% to over 97% of participants completing their first session.
Hardware Modalities (Headsets): Multimodal data is necessary; the best commercial multimodal headsets were rejected as inferior.
Custom headsets were fabricated by combining the best single-modality sensors and 3D-printed parts, designed to fit into a standardized Zarr 3 data format, allowing interchangeable use of sensors from multiple providers.
Headset design for training prioritizes sensor-density and expense tolerance; inference headsets (the final product) should only be finalized after model training reveals the minimum necessary sensor configuration.
Noise Reduction Trade-offs (Operational Efficiency): Noise reduction was constrained by the need for fast participant turnover (2–3 minutes between sessions).
Gel Rejection: Traditional EEG gel application was rejected as it takes 30 minutes, which would have more than doubled the marginal cost per hour of data. High-quality, spring-loaded dry electrodes were used instead.
Electrical Noise: Initial efforts to filter AC power failed. The functional, though operationally cumbersome, solution was relying solely on large (30lb) Anker DC batteries and physically turning off AC power to the building section.
Scaling Insight (Data Quantity Over Quality): After acquiring approximately 4,000–5,000 hours of data, the value of extreme noise reduction rapidly diminished.
The model capacity, trained on highly diverse, messy data, gained robustness and generalization, learning to distinguish signal from environment noise (e.g., room variations). This allowed the operation to cease inefficient noise-reduction practices (like continuous DC operation).
Operational Scaling and Cost Reduction: Marginal cost per usable hour was cut by ~40% between May and October.
Booking System: A custom booking suite was implemented to manage high volume (20 hours/day, 7 days/week), featuring dynamic pricing (to fill undesirable slots like 7 am Sunday) and dynamic overbooking (to account for no-shows).
Participant Management: A strict cap of 10 sessions per individual was introduced to prioritize dataset breadth (unique individuals) over schedule density (returning participants).
Critical Improvement: In August, the data collection backend was entirely rewritten using Zarr 3 to enable live, streaming sanity-checking of data (checking for modality dropout, timestamp drift, etc.). This single change cut marginal cost by ~30% and improved usable data collected by ~1.5x.
Supervision: EVERSECU cameras and a unified backend allow session managers to efficiently oversee and support multiple concurrent sessions without leaving the central station.
20:10 Current Status: The focus has shifted almost entirely away from data collection logistics (Step 1 complete) to model training (Step 2). The facility/dataset is available for external collaborations (contact@condu.it).