# Reproduction: OSF: On Pre-training and Scaling of Sleep Foundation Models

## Pages

| Page |
| --- |
| [Executive summary](#/executive-summary) |
| [Claim 1: SleepBench dataset scale: 166,500 hours across 9 NSRR sources and 21,000+ studies](#/claim-1-sleepbench-dataset-scale-166-500-hours-across-9-nsrr-sources-and-21-000-studies) |
| [Claim 2: MROS AUC: OSF 97.3 staging / 92.8 arousal vs SleepFM 96.4/90.3 and ViT 96.9/88.4 (Table 2)](#/claim-2-mros-auc-osf-97-3-staging-92-8-arousal-vs-sleepfm-96-4-90-3-and-vit-96-9-88-4-table-2) |
| [Claim 3: Ablation: time+channel masking 94.8 to 96.7 (Table 1)](#/claim-3-ablation-time-channel-masking-94-8-to-96-7-table-1) |
| [Claim 4: Missing-channel generalization: OSF vs SleepFM across 4 scenarios (Table 6)](#/claim-4-missing-channel-generalization-osf-vs-sleepfm-across-4-scenarios-table-6) |
| [Claim 5: Scaling with data fraction (1-100%) and encoder capacity (1M-85M)](#/claim-5-scaling-with-data-fraction-1-100-and-encoder-capacity-1m-85m) |
| [Conclusion](#/conclusion) |