| # Additional context-scaling experiments | |
| This directory provides **C3, C6, and C10** configurations using three, six, and ten reference RGB inputs. Each configuration contains the same **1,500 scene–query–target examples** as the [official benchmark](../README.md): **4,500 configuration records**, not 4,500 distinct test examples. | |
| | File | Purpose | | |
| |---|---| | |
| | `metadata.jsonl` | Ordered inputs and instructions for all three configurations | | |
| | `pairs.jsonl` | Correspondence between each official sample and its three configuration records | | |
| | `protocol.json` | Comparison and pairing rules | | |
| Group records by `shot`. Context lists are nested; preserve their frozen order and use consistent generation settings. Report each configuration separately, without pooling them into the official test denominator. | |
| Assets are shared with the repository-wide archives. VGGT scoring retains the same **ten reference views**, plus target and prediction, across all configurations. Set `VGGT_CANONICAL_JSON` to this directory's `metadata.jsonl` in the current project runtime; its ten-view records supply the scoring references. Held-out targets and scoring-only inputs must remain outside the generator context. | |
| Use the `context_scaling` Hugging Face configuration with `split="test"` and the same pinned revision as the downloaded assets. See the [main card](../README.md) for download, verification, extraction, and [data terms](../DATA_LICENSE.md). | |