|
Download ocr-tech/outputs/newa-balanced-sampling-v1/PROGRESS.md from prosperious/ocr-projects-archive: direct link, hf CLI and curl.
- Browser
- Download file 1.28 kB
-
https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/outputs/newa-balanced-sampling-v1/PROGRESS.md
- Command line
-
hf download hf://prosperious/ocr-projects-archive/ocr-tech/outputs/newa-balanced-sampling-v1/PROGRESS.md
-
curl -L -o PROGRESS.md https://huggingface.co/prosperious/ocr-projects-archive/resolve/main/ocr-tech/outputs/newa-balanced-sampling-v1/PROGRESS.md
1.28 kB
Newa balanced character sampling progress
Date: 2026-07-15
Status: DETERMINISTIC TRAIN-ONLY CANDIDATE; NOT ENABLED
The exact CTC200 train-only capsule was audited without extraction. This candidate leaves the capsule, Hitopadesa checkpoint-selection role, and Cambridge evaluation artifacts unchanged.
Measured source inventory:
- training rows:
1961; - Newa codepoint tokens/types:
146153/69; - extended-grapheme tokens/types:
87777/585; - excluded singleton codepoints/graphemes:
2/71.
Candidate schedule:
- scheduled rows:
2451(490additional references); - multiplicity histogram:
{'1': 1688, '2': 56, '3': 217}; - codepoint/grapheme features at the 32-exposure target:
13/42; - exact and 16-codepoint train/selection overlaps:
0/0; - novel image/text pairs:
0; checkpoint-selection rows copied:0.
The two singleton codepoints and 71 singleton graphemes are measured but intentionally not amplified pending label review. Cambridge labels are absent by design, so final corrected labels still require the independent post-review exact and 16-codepoint leakage audit. Training from this schedule would require a new payload/config/log binding; it is not authorized by the current frozen capsule or GPU approval gate.