Gespeichert in:
| Hauptverfasser: | , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2603.00173 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917300430438400 |
|---|---|
| author | Ryu, Simo Han, Chunghwan |
| author_facet | Ryu, Simo Han, Chunghwan |
| contents | We describe our experience training Summer-22B, a video foundation model developed from scratch. This report documents the engineering challenges, design decisions, and lessons learned while scaling from raw footage collection to a functional model trained on approximately 50 million clips. We outline our approach combining metadata-driven dataset curation, multi-stage filtering, $μ$P parameterization, and hypersphere-constrained optimization. We developed the Lavender Data system for dataset management and adopted inference-aware architectural choices. We share observations on what worked in our setting: dataset engineering consumed the majority of effort, architectural variants showed smaller differences than we expected, and $μ$P hyperparameter transfer appeared effective even under geometric constraints. We hope this account proves useful to others undertaking similar projects. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_00173 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Summer-22B: A Systematic Approach to Dataset Engineering and Training at Scale for Video Foundation Model Ryu, Simo Han, Chunghwan Computer Vision and Pattern Recognition Artificial Intelligence Machine Learning We describe our experience training Summer-22B, a video foundation model developed from scratch. This report documents the engineering challenges, design decisions, and lessons learned while scaling from raw footage collection to a functional model trained on approximately 50 million clips. We outline our approach combining metadata-driven dataset curation, multi-stage filtering, $μ$P parameterization, and hypersphere-constrained optimization. We developed the Lavender Data system for dataset management and adopted inference-aware architectural choices. We share observations on what worked in our setting: dataset engineering consumed the majority of effort, architectural variants showed smaller differences than we expected, and $μ$P hyperparameter transfer appeared effective even under geometric constraints. We hope this account proves useful to others undertaking similar projects. |
| title | Summer-22B: A Systematic Approach to Dataset Engineering and Training at Scale for Video Foundation Model |
| topic | Computer Vision and Pattern Recognition Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2603.00173 |