Jinfa commited on
Commit
d6949e1
·
verified ·
1 Parent(s): 007ec06

Initialize WorldFold (v7-enhance++) cache & checkpoint dataset card

Browse files
Files changed (1) hide show
  1. README.md +49 -0
README.md ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - video-text-to-text
5
+ - question-answering
6
+ tags:
7
+ - world-model
8
+ - matryoshka-representation
9
+ - video-spatial-reasoning
10
+ - seva
11
+ - qwen3-vl
12
+ ---
13
+
14
+ # WorldFold (`v7-enhance++`) Precomputed SEVA World-Model Feature Caches & Checkpoint
15
+
16
+ Official precomputed **Stable Virtual Camera (SEVA)** diffusion rollout feature caches (`[4, 2, 16, 4096]` `bfloat16` encoded via frozen `Qwen3-VL-8B-Instruct` visual tower) and trained **`Worldfold-v7-enhance++.pt`** adapter weights for **[WorldFold (`v7-enhance-pp`)](https://github.com/inFaaa/WorldFold/tree/v7-enhance-pp)**.
17
+
18
+ ---
19
+
20
+ ## 📦 Included Assets
21
+
22
+ | File / Directory | Coverage | Size | Target Extraction Path (`storage/outputs/`) |
23
+ |---|---|---:|---|
24
+ | `checkpoints/Worldfold-v7-enhance++.pt` | 2-Epoch `fp32` trained adapter checkpoint (`10,169` QA / `3,449` scenes) | `75.5 MB` | `nested_prior_runs/spec_v7_runs/checkpoints/Worldfold-v7-enhance++.pt` |
25
+ | `vsi_eval_seva_288scenes.tar.gz` | **VSI-Bench** (`288 / 288` test scenes, `5,130` questions) | `~600 MB` (`2.4 GB` raw) | `cache/vsi_eval_cache/scene_feats_real_seva/*.pt` |
26
+ | `ovos_eval_seva_339videos.tar.gz` | **OVO-S-Bench** (`339 / 339` test videos, `1,695` questions) | `~350 MB` (`1.4 GB` raw) | `cache/ovos_eval_cache/real_seva_feats/*.pt` |
27
+ | `mmsi_eval_seva_1106questions.tar.gz` | **MMSI-Video-Bench** (`1,106 / 1,106` test questions) | `~1.1 GB` (`4.4 GB` raw) | `cache/mmsi_eval_cache/real_seva_feats/question_*.pt` |
28
+ | `vsi_train10k_seva_only_3449scenes.pt` | **VSI-Train-10k** SEVA-only features (`3,449` training scenes: `wm_4x2`, `real_4x2`, `anchor_context`, `real_context`) | `~3.6 GB` | `cache/vsi_train10k_seva_only_3449scenes.pt` |
29
+ | `vsi_train10k_v7_unified/shard_00.pt` .. `shard_07.pt` | **VSI-Train-10k** full unified cache (`3,449` scenes including `vis_embeds_256` + 3-layer `deepstack_256` + `wm_4x2`) | `56.0 GB` | `cache/vsi_train10k_v7_unified/shard_00.pt` .. `shard_07.pt` |
30
+
31
+ ---
32
+
33
+ ## 🚀 Quick Usage
34
+
35
+ ```bash
36
+ # Download checkpoint and 3 benchmark SEVA caches (~2.1 GB total)
37
+ hf download Jinfa/WorldFold-v7-Cache \
38
+ checkpoints/Worldfold-v7-enhance++.pt \
39
+ vsi_eval_seva_288scenes.tar.gz \
40
+ ovos_eval_seva_339videos.tar.gz \
41
+ mmsi_eval_seva_1106questions.tar.gz \
42
+ --repo-type dataset --local-dir ./worldfold_cache
43
+
44
+ # Unpack into storage/outputs/cache/
45
+ mkdir -p storage/outputs/cache/vsi_eval_cache storage/outputs/cache/ovos_eval_cache storage/outputs/cache/mmsi_eval_cache
46
+ tar -xzf ./worldfold_cache/vsi_eval_seva_288scenes.tar.gz -C storage/outputs/cache/vsi_eval_cache/
47
+ tar -xzf ./worldfold_cache/ovos_eval_seva_339videos.tar.gz -C storage/outputs/cache/ovos_eval_cache/
48
+ tar -xzf ./worldfold_cache/mmsi_eval_seva_1106questions.tar.gz -C storage/outputs/cache/mmsi_eval_cache/
49
+ ```