Glyph scaling transfer — exact-image deduplication

Upload in progress for scaling training on another cluster.

Current release

• Input:1,713,825 historical native-quality AnyText candidate records.
• Output:1,459,580 unique images;254,245 duplicate image copies removed.
• Grouping:SHA256 of complete decoded512×512 RGB images. No center-crop/pHash/near-duplicate grouping.
• Per latest user instruction,no additional test-overlap or private-heldout exclusion is applied. Historical1,376,652 cluster-selection count is NOT this release's size.
• Different full annotations/captions for the same image are preserved as variants;originalsourceUIDs remain in aliases. These are annotation variants,not additional images. Sample by parent image and handle variants explicitly.
• Image bytes are unchanged. No resizing,recompression,reboxing,rewritten OCR labels or invented captions.

Files

anytext_unique/shards/*.tar: image bytes and pairedJSON. JSON contains imageSHA256,RGBSHA256,full annotation variants,caption,validity flags andsourceUID aliases.

anytext_unique/preparation.json: verified unique image/variant/input-record counts.

anytext_unique/shard_index.json: SHA256 and counts for everyuploaded shard. Wait for complete:true before using the fullrelease.

transfer/restore_unique.py: restore a downloaded snapshot to images/ and parents.jsonl,without dropping variants. The data records use portable relative filenames. Old transfer/restore.py targets a superseded export plan and must not be used for thisrelease.

Any old100K/R/Q/D selections are subsets of these images,not extra corpora. Cross-source overlaps should remove augmentation entries while preservingAnyText. Source-specific terms remain applicable;this collection does not replace original licenses.

Augmentation

The previous two-pass augmentation pool contains71,882 net-new parents across10sources. It is not a fully corrected release:the audit found missing clarity/ignore-overlap checks,and the corrected version currently has60preview parents/83crops. Augmentation upload remains separately pending;do not silently label the oldfullpool as corrected. Details are in transfer/inventory.json.

No model weights,optimizer states,raw archives or experimental generated images are uploaded. Existing training continues independently.