Polish SGOCR v1 release card

README.md

@@ -13,129 +13,104 @@ tags:
 - metadata-only
 size_categories:
 - 1K<n<10K
-pretty_name: SGOCR Mixed 3000 Balanced DAM01 Primary Gemma
+pretty_name: SGOCR
 ---
 
-# SGOCR Mixed 3000 Balanced DAM01 Primary Gemma
+# SGOCR
 
-This is a metadata-only release of `sgocr_mixed3000_balanced_dam01_20260417_173000_primary_gemma`. The repository does **not** host upstream PNG/JPG image files. Each row references source-specific image IDs and source paths so users can join against their own local copies of ChartQA, TextOCR, and COCO/COCO-Text.
+SGOCR is a source-referenced OCR visual question answering dataset. It contains synthetic question-answer pairs grounded in text regions and visual anchors from ChartQA, TextOCR, and COCO/COCO-Text images.
 
-The included train split is the completed primary Gemma-anchor SGOCR output. The canceled partial second `final_teacher` pass is not included.
+This repository is intentionally metadata-first: it does not redistribute the full upstream image corpora. Each sample includes source-specific image identifiers so users can join against their own licensed copies of the source datasets.
 
-## Files
-
-- `data/train.jsonl`: flattened accepted QA rows, one row per question-answer sample.
-
-- `data/source_images.jsonl`: one row per source image referenced by accepted QA rows.
-
-- `metadata/accepted_dataset_original.jsonl.gz`: compressed original accepted rows as emitted by the run.
-
-- `metadata/rejected_dataset.jsonl.gz` and `metadata/raw_results.jsonl.gz`: compressed audit artifacts from generation.
-
-- `metadata/summary.json`, `metadata/source_manifest.json`, `metadata/run_report.json`: run metadata and provenance.
-
-## Counts
-
-- Accepted QA rows: `5737`
-
-- Referenced source images: `1958`
-
-- Raw teacher rows: `15214`
-
-- Rejected rows: `9477`
-
-## Source Breakdown
-
-- `chartqa`: `1964` QA rows, `754` images
-
-- `coco_text`: `1987` QA rows, `570` images
-
-- `textocr`: `1786` QA rows, `634` images
-
-## Question Type Breakdown
-
-- `DIRECT_READ`: `2778` rows
-
-- `REVERSE_GROUND`: `409` rows
+- Project repo: https://github.com/cothogonal/sgocr-dataset-pipeline
+- Current release tag: `v1.0.0`
+- Hugging Face revisions: this dataset repo is Git-backed. Future releases can add rows or columns while preserving this snapshot through tags and commit history.
 
-- `TEXT_PROPERTY`: `915` rows
+## Dataset Snapshot
 
-- `YES_NO`: `1635` rows
+| Split | QA rows | Source images | Image hosting |
+|---|---:|---:|---|
+| `train` | 5,737 | 1,958 | Source-referenced; full image corpus not redistributed |
 
-## Quality Tier Breakdown
+## Source Mix
 
-- `tier_a`: `3316` rows
+| Source | QA rows | Source images | Join key |
+|---|---:|---:|---|
+| `chartqa` | 1,964 | 754 | `source_image_id` = ChartQA image stem |
+| `textocr` | 1,786 | 634 | `source_image_id` = TextOCR image id |
+| `coco_text` | 1,987 | 570 | `source_image_id` = COCO train2014 numeric id |
 
-- `tier_b`: `2418` rows
+## Question Types
 
-- `unscored`: `3` rows
+| Question type | Rows | Description |
+|---|---:|---|
+| `DIRECT_READ` | 2,778 | Read visible text in a grounded region. |
+| `REVERSE_GROUND` | 409 | Identify or localize an anchor from text-region context. |
+| `TEXT_PROPERTY` | 915 | Ask about properties of visible text, such as count or case. |
+| `YES_NO` | 1,635 | Verify whether a grounded region contains a proposed text/value. |
 
-## Example Rows
+## Example Samples
 
-### Example 1
+### Example 1: Direct Read
 
-- Source: `coco_text` / `train` / `282357`
+![Example 1](assets/examples/example_1.jpg)
 
-- Question type: `DIRECT_READ`
+| Field | Value |
+|---|---|
+| Source | `coco_text` / `train` / `282357` |
+| Question | What does the sign say, specifically the upper-left text around the top-center area of the image? |
+| Answer | `BHAR` |
+| Anchor | `sign` |
+| Relation | `above` |
 
-- Question: What text is on the face, specifically the upper-right text near the top-center area of the image?
+### Example 2: Yes No
 
-- Answer: `AQUARIUM`
+![Example 2](assets/examples/example_2.png)
 
-- Anchor: `face`; reference: `None`; relation: `above`
+| Field | Value |
+|---|---|
+| Source | `chartqa` / `shared` / `ce1423c017bd7c98ff8c9414f03ac33080d0036910639323ec09b7a413ee5b8c` |
+| Question | Does the black label's upper-left text near the upper-left area of the image say '25-34'? |
+| Answer | `No` |
+| Anchor | `blue bar` |
+| Relation | `on` |
 
-- Source path hint: `data/vm_ssl/raw/coco_text_materialized/train/COCO_train2014_000000282357.jpg`
-
-### Example 2
-
-- Source: `chartqa` / `shared` / `ce1423c017bd7c98ff8c9414f03ac33080d0036910639323ec09b7a413ee5b8c`
-
-- Question type: `YES_NO`
-
-- Question: Does the black label's upper-left text near the upper-left area of the image say '25-34'?
-
-- Answer: `No`
-
-- Anchor: `blue bar`; reference: `None`; relation: `on`
-
-- Source path hint: `data/vm_ssl/raw/chartqa/images/ce/ce1423c017bd7c98ff8c9414f03ac33080d0036910639323ec09b7a413ee5b8c.png`
-
-## How To Join Images
-
-This repo intentionally does not redistribute images. Download the upstream datasets yourself, then join using `source_dataset`, `source_split`, and `source_image_id`.
-
-- `chartqa`: download ChartQA from its official distribution. Join `source_image_id` to the image file stem, usually `<source_image_id>.png`; the original local path hint is in `source_image_path`.
-
-- `textocr`: download TextOCR v0.1 train data. Join `source_image_id` to `TextOCR_0.1_train.json["imgs"][source_image_id]["file_name"]`.
-
-- `coco_text`: download COCO train2014 images and COCO-Text metadata if needed. Join `source_image_id` to `COCO_train2014_<12-digit-source_image_id>.jpg`.
-
-## Field Guide
-
-### Identity and source columns
-
-`sample_id`, `image_id`, `source_dataset`, `source_split`, `source_image_id`, `source_image_path`, `source_file_name`, `upstream_dataset`, `join_hint`, `dataset_source`
-
-### QA columns
-
-`question`, `answer`, `question_type`, `answer_level`, `answer_type`, `answer_source`, `text_length_bin`, `text_case`
+## Files
 
-### Grounding columns
+| Path | Purpose |
+|---|---|
+| `data/train.jsonl` | Main training/evaluation table, one QA sample per row. |
+| `data/source_images.jsonl` | Image-level join table with one row per referenced source image. |
+| `metadata/summary.json` | Compact generation summary and aggregate statistics. |
+| `metadata/source_manifest.json` | Source-image manifest for the sampled image universe. |
+| `metadata/run_report.json` | Provenance metadata for this release. |
+| `metadata/*.jsonl.gz` | Compressed audit artifacts for users who need generation traces. |
 
-`anchor_label`, `ref_label`, `relation`, `ann_id`, `consensus_tier`, `text_node_ids`, `text_polygon`, `text_bbox_xywh_*`, `anchor_xyxy_*`, `ref_xyxy_*`
+## Main Columns
 
-### Quality/provenance columns
+| Group | Columns |
+|---|---|
+| Identity | `sample_id`, `image_id` |
+| Source join | `source_dataset`, `source_split`, `source_image_id`, `source_file_name`, `upstream_dataset`, `join_hint` |
+| QA | `question`, `answer`, `question_type`, `answer_level`, `answer_type`, `answer_source` |
+| Text and anchors | `anchor_label`, `ref_label`, `relation`, `text_node_ids`, `text_polygon`, `text_bbox_xywh_*`, `anchor_xyxy_*`, `ref_xyxy_*` |
+| Image metadata | `image_width`, `image_height` |
+| Quality/provenance | `ocr_confidence`, `resolvable`, `unique`, `quality_tier`, `inline_frontier_correct`, `teacher_provider`, `teacher_model`, `prompt_variant` |
 
-`ocr_confidence`, `resolvable`, `unique`, `quality_tier`, `inline_frontier_correct`, `teacher_provider`, `teacher_model`, `prompt_variant`, `neighboring_text_count`, `resolvability_reason`
+## Joining Images
 
-## Notes
+| Source | How to obtain images | Join procedure |
+|---|---|---|
+| ChartQA | Download ChartQA from its official distribution. | Join `source_image_id` to the ChartQA image stem, typically `<source_image_id>.png`. |
+| TextOCR | Download TextOCR v0.1 train data. | Join `source_image_id` through `TextOCR_0.1_train.json["imgs"][source_image_id]["file_name"]`. |
+| COCO / COCO-Text | Download COCO train2014 images and COCO-Text metadata as needed. | Join `source_image_id` to `COCO_train2014_<12-digit-source_image_id>.jpg`. |
 
-- `inline_frontier_correct` is retained as provenance from the completed primary run. It came from the source artifact and should not be interpreted as a fresh benchmark in this release.
+## Versioning
 
-- Box suffixes ending in `_x2_or_w` / `_y2_or_h` preserve the original coordinate convention: text boxes are xywh; anchor/ref boxes are xyxy. The array-valued box column is included alongside split numeric fields.
+Hugging Face dataset repositories are Git-backed. Use the `v1.0.0` tag for this release. Future SGOCR releases can be published as additional commits and tags in the same `SGOCR` repository, keeping old snapshots addressable by tag or commit SHA.
 
-- Review upstream dataset licenses and terms before using or redistributing joined images.
+## License And Source Dataset Terms
 
-## Storage Note
+This repository contains SGOCR annotations and metadata plus two small illustrative example images in the dataset card. The full upstream image corpora are not redistributed here. Users are responsible for obtaining ChartQA, TextOCR, COCO, and COCO-Text under their respective licenses and terms of use before joining images to the metadata.
 
-Large audit JSONL artifacts are gzip-compressed. The main `data/train.jsonl` and `data/source_images.jsonl` remain plain JSONL for easy loading and dataset preview.
+The SGOCR annotations are released for research and dataset-development use. Because samples are derived from third-party image datasets, downstream redistribution or commercial use may require compliance with the upstream dataset licenses in addition to this repository's terms.