| license: apache-2.0 | |
| base_model: ThinkMorph/ThinkMorph-7B | |
| tags: | |
| - vision-language | |
| - jsd | |
| - imagination-audit | |
| # Spatial-UMM (JSD L0 Visual Search, seed 3) | |
| This repository is the EMA checkpoint of **JSD L0 VS s3**, the highest point estimate in the current same-judge 9-benchmark series. | |
| | Item | Value | | |
| |---|---| | |
| | Init | ThinkMorph-7B | | |
| | Objective | Jensen-Shannon distillation, understanding side (`visual_gen=False`) | | |
| | Data | Imagination-Audit pass set, L0 Visual Search, train only (GT action) | | |
| | Seed | 4399 | | |
| | Checkpoint | step 999 EMA | | |
| | Same-judge Avg(9) | **64.28** (local Qwen3-30B-A3B, und-decode) | | |
| `model.safetensors` is that EMA. Tokenizer, ViT/LLM configs, and `ae.safetensors` are the ThinkMorph-7B companions required to load the student. | |
| The 64.28 score is the best point estimate under this protocol. A paired bootstrap against the original L0 VS run (62.91) has a 95% interval that includes 0, so this checkpoint is the selected peak, not a claim of a significant gain over that earlier seed. It does beat ThinkMorph, the length-matched JSD run, and the nav+VS mix where those intervals exclude 0. | |
| Do not compare this number to the published Visual-OPSD average of 74.03. That figure uses a different judge. | |