SEELE-AI/EVA-Edit

EVA01-Edit dataset

This release contains the mesh and prompt records used for the EVA01-Edit data. The recommended Hugging Face distribution uses uncompressed TAR shards so that complete edit groups can be streamed without creating hundreds of thousands of independent repository objects.

Scope

The public release has two modalities: 3D meshes and text prompts. It contains only:

• GLB mesh files for the complete three-stage edit chains.
• The base caption and edit instruction associated with each stage.
• Relative paths for mesh lookup and prompt metadata.
• Release metadata checksums.

Point clouds, latent tensors, latent metadata, optimizer state, trainer state, and other training artifacts are excluded from this release.

Dataset structure

Each complete group contains the following ordered stages:

1. image0_base: the base mesh and its base caption.
2. image1_edit1: the first edited mesh and the instruction applied to the base stage.
3. image2_edit2: the second edited mesh and the instruction applied to the first edited stage.

The package provides two manifest views:

• manifest/eva01_edit_three_stage.jsonl contains one record per complete group and preserves the full three-stage chain.
• manifest/eva01_edit_two_stage.jsonl contains one record per adjacent edit transition. Each group contributes a base-to-first-edit pair and a first-edit-to-second-edit pair.

The release contains 119,905 complete groups, 359,715 stage records, and 239,810 adjacent edit pairs.

The relative mesh layout inside each shard is:

meshes/<group_key>/<stage>/<filename>.glb

Hugging Face distribution format

The upload package should use the following layout:

README.md
LICENSE
manifest/eva01_edit_three_stage.jsonl
manifest/eva01_edit_two_stage.jsonl
manifest/eva01_edit_three_stage_sharded.jsonl
manifest/eva01_edit_shard_index.jsonl
shards/shard-000000.tar
shards/shard-000001.tar
...

Each shard is an uncompressed TAR archive with a target size of approximately 2 GiB. A complete three-stage group is kept in one shard; a group is never split across shards. The TAR member names are the same relative mesh paths shown above. manifest/eva01_edit_shard_index.jsonl maps each group and member to its shard, while manifest/eva01_edit_three_stage_sharded.jsonl adds the shard reference to every mesh record. The uncompressed format avoids wasting CPU on already-compressed GLB buffers and supports standard TAR streaming and partial downloads.

The shard packer accepts a mounted source tree and writes one shard at a time, so the full dataset does not need to fit on local disk.

A three-stage manifest record has the following shape:

{
  "schema_version": "eva01_edit",
  "release_version": "eva01_edit",
  "group_key": "group_...",
  "stage_chain": ["image0_base", "image1_edit1", "image2_edit2"],
  "stages": [
    {
      "stage": "image0_base",
      "mesh": {
        "path": "meshes/group_.../image0_base/mesh_....glb",
        "filename": "mesh_....glb"
      },
      "prompt": {
        "text": "...",
        "kind": "base_caption",
        "parent_stage": null
      }
    }
  ]
}

Packing upload shards

For the full release, pack the sanitized three-stage manifest into approximately 2 GiB TAR shards:

python tools/pack_shards_from_bos.py \
  --manifest manifest/eva01_edit_three_stage.jsonl \
  --source-root /path/to/source/meshes \
  --output-root .

When an input manifest contains source checksums, the packer verifies every source GLB while writing the TAR member. The public output records only TAR shard checksums. It writes shards/, manifest/eva01_edit_three_stage_sharded.jsonl, and manifest/eva01_edit_shard_index.jsonl; this release does not publish shard checksums. Use --max-groups N for a bounded smoke test. A production run can be executed on a host with enough temporary space for one shard and uploaded shard by shard to the Hugging Face repository.

Materializing mesh files

The manifests are metadata-only until the GLB files are materialized into the meshes/ tree. The release tool can read from a mounted source dataset and copy one group or a bounded batch:

python tools/materialize_from_bos.py \
  --manifest manifest/eva01_edit_three_stage.jsonl \
  --source-root /path/to/source/meshes \
  --output-root . \
  --group-key group_...

Use --max-groups N for a bounded batch. Use --verify-only to check source availability and SHA-256 values without writing files. The source root is intentionally supplied by the user and is not embedded in the public manifest.

Verification

The public manifest does not include a SHA-256 field for every mesh. The public release does not publish per-mesh or TAR checksums. Hugging Face upload integrity and the manifest structure provide the release verification surface. The TAR member names and prompt records are listed in the corresponding manifest. The packer verifies source meshes when input checksums are available, while per-mesh checksums are intentionally omitted from the public schema.

Attribution and limitations

The meshes are released as the EVA01-Edit mesh and prompt data associated with EVA01. The prompts describe the base object and the sequential edit instructions; they are not human-authored annotations for every geometric property. The GLB files may contain vertex colors and material information supported by the source generation pipeline. Consumers should validate rendering and downstream licensing requirements for their intended use.

This release does not include the EVA01 model weights or the TRELLIS.2 VAE/decoder weights. Those assets are distributed separately by their respective model repositories.

License

This dataset is released under the Apache License 2.0. See LICENSE. The license applies to the released dataset materials; users remain responsible for checking any rights or restrictions associated with downstream assets and intended use.

Citation

@misc{eva01_2026,
  title  = {EVA01: Unified Native 3D Understanding and Generation via Mixture-of-Transformers},
  author = {Zongyuan Yang and Mingjing Yi and Wanli Ma and Chenzhuo Fan and Bocheng Li and Baolin Liu and Yuke Lou and Yingde Song and Yongping Xiong and Zhengdong Guo and Shimu Wang},
  year   = {2026},
  eprint = {2605.16745},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV}
}