Clean public benchmark docs

hf_eval_results/HF_UPSTREAM_REQUEST.md

@@ -1,92 +0,0 @@
-# Requesting `real5-omnidocbench` Support Upstream
-
-To make the Hugging Face native Benchmark UI recognize this benchmark, Hugging Face needs to add a new evaluation framework identifier:
-
-```text
-real5-omnidocbench
-```
-
-The relevant upstream enum is currently in:
-
-```text
-https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/eval.ts
-```
-
-## Proposed Enum Entry
-
-```ts
-"real5-omnidocbench": {
-  name: "real5-omnidocbench",
-  description:
-    "Real5-OmniDocBench is a real-world document parsing robustness benchmark covering five physical acquisition scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew.",
-  url: "https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench",
-},
-```
-
-## Dataset Configuration
-
-The dataset repository already contains a root `eval.yaml` using the requested framework:
-
-```yaml
-name: Real5-OmniDocBench
-description: >
-  Real5-OmniDocBench is a real-world document parsing robustness benchmark
-  covering five physical acquisition scenarios: Scanning, Warping,
-  Screen-Photography, Illumination, and Skew.
-evaluation_framework: real5-omnidocbench
-
-tasks:
-  - id: overall
-    config: default
-    split: test
-  - id: scanning
-    config: default
-    split: test
-  - id: warping
-    config: default
-    split: test
-  - id: screen_photography
-    config: default
-    split: test
-  - id: illumination
-    config: default
-    split: test
-  - id: skew
-    config: default
-    split: test
-```
-
-## Message Template
-
-```text
-Hi Hugging Face team,
-
-We would like to add a new Hub Benchmark evaluation framework:
-
-real5-omnidocbench
-
-Dataset:
-https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench
-
-Paper:
-https://arxiv.org/abs/2603.04205
-
-Reason:
-Real5-OmniDocBench evaluates document parsing robustness under five real-world acquisition scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. Existing framework identifiers such as olmocr-bench, mdpbench, and parsebench do not identify this benchmark directly.
-
-Proposed enum entry:
-
-"real5-omnidocbench": {
-  name: "real5-omnidocbench",
-  description:
-    "Real5-OmniDocBench is a real-world document parsing robustness benchmark covering five physical acquisition scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew.",
-  url: "https://huggingface.co/datasets/PaddlePaddle/Real5-OmniDocBench",
-}
-
-The dataset root eval.yaml already uses:
-evaluation_framework: real5-omnidocbench
-
-Could you add this framework identifier to the supported enum and add the dataset to the Benchmark allow-list so the native Benchmark Leaderboard UI can appear?
-
-Thanks.
-```

hf_eval_results/README.md

@@ -1,34 +1,7 @@
-# Hugging Face Eval Results Support
+# Hugging Face Eval Results
 
-This directory prepares Real5-OmniDocBench for Hugging Face's model-side
-`.eval_results/*.yaml` mechanism.
-
-## Current Support Status
-
-Not fully enabled on the Hub yet.
-
-Hugging Face Evaluation Results require:
-
-1. The dataset repo to have a root `eval.yaml`.
-2. The `evaluation_framework` in that file to be one of Hugging Face's supported framework identifiers.
-3. The benchmark dataset to be accepted by the beta allow-list.
-4. Model repos to submit `.eval_results/*.yaml` files that reference this dataset and task IDs.
-
-As of July 11, 2026, Hugging Face documents this feature as work in progress and benchmark registration as beta / allow-listed. The supported framework enum currently includes document-related frameworks such as `olmocr-bench`, `mdpbench`, and `parsebench`, but not `real5-omnidocbench`.
-
-The intended framework identifier for this benchmark is:
-
-```yaml
-evaluation_framework: real5-omnidocbench
-```
-
-Hugging Face must add this identifier to its supported evaluation framework enum before the Hub can fully recognize the dataset as a native benchmark.
-
-For that reason, the repository includes the target framework name explicitly so the upstream request is unambiguous. If the Hub rejects or ignores the framework before upstream support lands, keep the file as the desired target state and reference it in the Hugging Face support request.
-
-## Dataset-Side Candidate
-
-`eval.yaml.candidate` defines six Hub leaderboard tasks:
+Real5-OmniDocBench defines six Hub leaderboard tasks in the repository root
+`eval.yaml`:
 
 - `overall`
 - `scanning`
@@ -37,7 +10,8 @@ For that reason, the repository includes the target framework name explicitly so
 - `illumination`
 - `skew`
 
-These correspond to the six score columns in the README Overall table. All six are "higher is better" scores in the current benchmark presentation.
+These correspond to the primary scores shown in the benchmark leaderboard. All
+six are higher-is-better metrics.
 
 ## Model-Side Submission Format
 
@@ -66,7 +40,8 @@ Each YAML list item reports one task:
 
 ## Generate Files From Current README Data
 
-The static leaderboard data must already be generated from the external source file:
+The Space leaderboard data must already be generated from the external source
+file:
 
 ```bash
 python leaderboard/scripts/build_from_source_data.py
@@ -92,4 +67,7 @@ python hf_eval_results/scripts/validate_eval_results.py hf_eval_results/generate
 
 ## Important Limitation
 
-The Hub `.eval_results` schema stores one numeric `value` per dataset task. It does not directly encode the full multi-metric rows from the README tables. This support therefore maps the Hub-native leaderboard to the six primary Overall-table scores. The richer scenario-level metrics such as `TextEdit`, `FormulaCDM`, `TableTEDS`, and `Reading OrderEdit` remain available in the Static Space leaderboard.
+The Hub `.eval_results` schema stores one numeric `value` per dataset task.
+This repository uses the six primary task scores for Hub-native submissions.
+The richer scenario-level metrics remain available in the Static Space
+leaderboard.

hf_eval_results/eval.yaml.candidate

@@ -1,26 +0,0 @@
-name: Real5-OmniDocBench
-description: >
-  Real5-OmniDocBench is a real-world document parsing robustness benchmark
-  covering five physical acquisition scenarios: Scanning, Warping,
-  Screen-Photography, Illumination, and Skew.
-evaluation_framework: real5-omnidocbench
-
-tasks:
-  - id: overall
-    config: default
-    split: test
-  - id: scanning
-    config: default
-    split: test
-  - id: warping
-    config: default
-    split: test
-  - id: screen_photography
-    config: default
-    split: test
-  - id: illumination
-    config: default
-    split: test
-  - id: skew
-    config: default
-    split: test

leaderboard/DEPLOY.md

@@ -41,14 +41,11 @@ Hugging Face also has a Hub-level Evaluation Results feature. In that flow:
 2. Each model repository publishes scores in `.eval_results/*.yaml`.
 3. Hugging Face aggregates the model results back onto the benchmark dataset page.
 
-This is better for decentralized community submissions, but it requires model repositories to publish result files and the benchmark feature is currently marked as work in progress / beta allow-listed in the official docs.
+This is useful for decentralized community submissions, and it requires model repositories to publish result files.
 
 For this benchmark, the current README already has complete published tables, so the Static Space is the lowest-friction path.
 
 The repository also includes `hf_eval_results/` with:
 
-- `eval.yaml.candidate` for the dataset-side benchmark registration.
 - `examples/model_eval_results.yaml` for model repositories.
 - scripts to generate and validate `.eval_results/*.yaml` files from the README-derived leaderboard data.
-
-Do not move `hf_eval_results/eval.yaml.candidate` to the dataset repository root as `eval.yaml` until Hugging Face confirms the benchmark allow-list and evaluation framework choice.

leaderboard/README.md

@@ -29,4 +29,4 @@ The generated `leaderboard/data/` directory is ignored by the dataset repository
 Hugging Face supports leaderboard-style presentation in two relevant ways:
 
 - Community managed leaderboards live on Spaces and are appropriate for benchmark pages with custom tables and filters.
-- Evaluation Results are a Hub-level benchmark feature where a dataset repo defines `eval.yaml` and model repos publish `.eval_results/*.yaml`; this is currently marked as work in progress and beta allow-listed.
+- Evaluation Results are a Hub-level benchmark feature where a dataset repo defines `eval.yaml` and model repos publish `.eval_results/*.yaml`.