laya-neuro (early checkpoint)

An unofficial, independent fine-tune of Laya by Nandhakishor M / Convai Innovations — not affiliated with or endorsed by the original project.

Early checkpoint, not the final result. Trained on 3,000 of the 17,820 available training examples for 2 epochs (a fast initial run). A full run on all training data is in progress. See Aditharavind/laya-neuro for the full pipeline and updated results.

A Laya checkpoint (ModernBERT-large encoder, 421M params) fine-tuned via RLCD to classify cognitive workload (high/low) from EEG-derived features, using the public STEW dataset (48 subjects, 14-channel Emotiv EPOC EEG, resting-baseline vs. SIMKAP-multitasking).

Evaluated on the honest, subject-disjoint protocol: the test set is 12 subjects held out entirely from training (STEW's own pre-verified subject-disjoint fold), so this number reflects generalization to a new person, not a same-subject shortcut.

Training

• Base: convaiinnovations/laya
• Data: Aditharavind/laya-neuro-decisions, cross_subject/ split, 3,000-example random subset of the 17,820-example train set (early run)
• Recipe: RLCD (GRPO-style policy gradient over proper scoring rules + soft cross-entropy), 2 epochs, single 6GB consumer GPU, 8-bit AdamW (bitsandbytes)

Results (this checkpoint)

Metric	Cross-subject test (n=7,128, 12 held-out subjects)
Accuracy	66.3%
ECE	8.3%
Brier score	0.449

For comparison, classical ML baselines trained on identical features and the identical subject-disjoint split: LDA 67.3%, SVM (RBF) 67.1%, Random Forest 65.7%. This early Laya checkpoint is already in the same range as those baselines despite seeing under 17% of the available training data.

Input format

{
  "state": {"delta_mean": 0.4, "theta_mean": 0.14, "alpha_mean": 0.08, "beta_mean": 0.11,
            "gamma_mean": 0.05, "frontal_theta": 0.14, "posterior_alpha": 0.11,
            "frontal_theta_posterior_alpha_ratio": 1.3, "engagement_index": 0.5,
            "theta_alpha_ratio": 1.7},
  "questions": {"high_workload": {"type": "noul",
                "instructions": "Given this subject's EEG-derived features for this 2-second window, is this subject currently experiencing high cognitive workload?"}}
}

Features are relative EEG band power (Welch PSD, 5 canonical bands) plus engineered ratios, extracted per 2-second epoch — see eeg/features.py in the GitHub repo for the exact extraction code.

License / attribution

Fine-tunes Laya (Nandhakishor M / Convai Innovations, Apache 2.0). Dataset: STEW (Lim, Sourina & Wang), via the monster-monash/STEW Hugging Face mirror, CC BY 4.0.