Oratis's picture
归档历史采集元数据与训练笔记,保留原始资格矛盾和固定视频出处
21fe504 verified
|
Raw History Blame Contribute Delete
5.02 kB
metadata
pretty_name: Diogenes Vision Language Research Archive v01
license: other
license_name: historical-research-and-upstream-data-terms
license_link: >-
  https://huggingface.co/datasets/DiogenesLab/Diogenes_Vision_Language_Research_Archive_v01/blob/main/NOTICE.md
tags:
  - archival
  - evaluation
  - vision-language
  - non-trainable
size_categories:
  - 10K<n<100K

Diogenes Vision–Language Research Archive v01

The project has closed. This owner-authorized archive preserves 466 research files / 34,127,466 public bytes: 452 historical reports/configurations, nine Qwen teacher-label files, four Gaming500-derived benchmark manifests, and one Gemini evaluation-only label file. It contains research text and structured results; source images, model weights, databases and operational journals/logs are outside this dataset.

Folder Contents Historical use limits
research/ Metrics, predictions, configurations and research notes Historical experiments; no new quality or reproducibility certification
teacher-labels/qwen/ Nine teacher-label sets with original teacher and hybrid-source attribution Original is_trainable=false values are preserved
benchmark-manifests/ Four Gaming500-derived sample indexes Original is_trainable=false; upstream noncommercial terms remain applicable
eval-only/gemini/ 300 Gemini 2.5 Flash evaluation labels Evaluation-only, not training data

The archived values are historical observations, including failures, inconsistencies, ablations and model-generated predictions that may be incorrect or hallucinated. No benchmark was rerun and no historical score was recalculated for this publication. Numerical values, booleans and container structure in JSON/JSONL files were preserved and checked. Present-day upstream licensing does not establish the exact dataset/model revision used by a historical run when its configuration did not record one.

Explicit public copies

104 files are documented privacy copies. They replace 389 private home-path occurrences, seven Windows user-path occurrences, one model-predicted email string and nine private-network address strings. Address/email redaction does not assert that a model prediction describes a real person or machine. Other files preserve their exact original bytes. Complete private originals remain backed up.

MANIFEST.jsonl maps every original relative path and original SHA-256/size to its public path and SHA-256/size. PUBLICATION-COPIES.jsonl lists each changed file and transformation counts. Raw private local/cloud paths are not included in either map. Existing checksums and identifiers within historical reports are not rewritten to certify changed text, and metrics referring to earlier original predictions remain historical rather than scores recomputed from redacted predictions.

Provenance and use

Gaming500-derived manifests and labels attribute Markov AI's Gaming 500 Hours, whose current card declares CC BY-NC 4.0. The current source revision observed during closeout was 0e84517002c33b030e0aff30c8afc7714a8fc8cb; this is not a claim that all historical runs used that revision. Source benchmark images are not redistributed here. Keep attribution and noncommercial restrictions when using the corresponding derived material.

Model families referenced by experiments include Qwen3-VL and DINOv2; their official source/model licenses are separate from rights in benchmark data or generated labels. Archived teacher identifiers and evaluation-only flags remain unchanged. The public Diogenes source snapshot documents available code and reconstruction limits. Read NOTICE.md before reuse.

Synthetic fixtures and additional historical code

Procedural smoke fixtures and eight historical code snapshots preserve 144 logical synthetic NPZ/PNG files with exact aliases, fixed annotation links, and the original experimental limitations.

Additional historical training source

Training launchers, experimental records, and modified LongLive code preserve 16 additional historical files with exact source/public hashes and the unchanged upstream license; scientific limitations and minimal privacy edits are documented.

Historical capture metadata and training notes

Raw-input capture metadata and historical research notes preserve 1,716 distinct payloads with exact source/public hashes and fixed video lineage. Historical training-status contradictions, original 16-fps dataset declarations, and minimal privacy edits are documented.