Datasets:

Modalities:
3D
ArXiv:
License:
next / README.md
YipengGao's picture
Add files using upload-large-folder tool
d976411 verified
|
Raw History Blame Contribute Delete
3.14 kB
metadata
license: cc-by-nc-sa-4.0
pretty_name: 3DCodeVerse-Next
tags:
  - 3d
  - code-generation
  - 3dcodeverse
extra_gated_heading: Access 3DCodeVerse-Next
extra_gated_prompt: >-
  3DCodeVerse-Next holds work-in-progress data and run logs from the 3DCodeVerse
  project, shared for non-commercial research (CC BY-NC-SA 4.0; per-source
  licenses also apply). Requests are reviewed by hand. The information you
  provide is visible to the dataset authors.
extra_gated_fields:
  Full name: text
  Affiliation (company / university): text
  Country: country
  Intended use: text
  I will use this data only for non-commercial research and will respect each source license: checkbox
  I will not redistribute this work-in-progress data without the authors' permission: checkbox
extra_gated_button_content: Request access

3DCodeVerse-Next

Staging area for the next release of 3DCodeVerse: new subsets, harness run logs and QC reports, uploaded here for review before they are merged into the main dataset.

Contents are work in progress and may change or be removed without notice. Only members of the 3dcodeverse organization can upload.

📄 Paper · 🌐 Project page · 💻 Code

Layout (2026-11-07 upload)

Folder Local origin What
harness/runs/ 3dcodeverse/harness/runs 3DCodeVerse harness runs (argon_w1/w2, calib, smoke)
3dcodeverse_runs/ ~/3dcodeverse_runs coverage / test runs, scratch archive, v01 review (2026-11-03)
demos/ ~/3dcodeverse_demos gemini_objects, gemini_scenes, omni_whitebox logs and runs
eval/logs/ 3dcodeverse_data/3dcodeverse_eval/logs eval queue / scoring logs
meta3dcode/runs/ ~/meta3dcode/runs meta-harness runs
astrafilm/runs/ ~/astrafilm-runs film runs (alphago-move78, interstellar, optimus, pooh)
projects/<name>/outputs/ ~/projects/*/outputs astra3d-brilliana, scene_multifile, scene_multifile_graphics outputs

Each first-level run folder is either raw files (small folders) or packed:

<run>/part-000.tar ...   plain tar, <= 2 GB each; symlinks kept as links
<run>/index.parquet      one row per file: path, size, tar, byte_start, is_link, link_target, mtime

Read one file without downloading the tar:

import pandas as pd, httpx
from huggingface_hub import hf_hub_download, get_token
run = "harness/runs/smoke"
idx = pd.read_parquet(hf_hub_download("3dcodeverse/next", f"{run}/index.parquet", repo_type="dataset"))
r = idx[idx.path.str.endswith("run.log")].iloc[0]
url = f"https://huggingface.co/datasets/3dcodeverse/next/resolve/main/{run}/{r.tar}"
data = httpx.get(url, headers={"Authorization": f"Bearer {get_token()}",
                 "Range": f"bytes={r.byte_start}-{r.byte_start + r.size - 1}"}, follow_redirects=True).content

Not uploaded: virtualenvs, node_modules, __pycache__, .git, CLI home dirs. A secret scan was run before upload.