Skip to content

Releases

Unlimited-OCR-2: the six cases' inputs and stage dumps (#2074 at b68ea38f2)

Choose a tag to compare

@BobClawblaw BobClawblaw released this 11 Nov 21:52

Inputs for Unlimited-OCR-2's six cases (ashhart#2074, head b68ea38f2): the five page images, each case's stage dumps from the release's own code (torch bf16 and float32, each with its top2 sidecar), and the two block-case files tf-uocr-blocks reads. SHA256SUMS covers every file.

Pages (the same bytes as release uocr-0af3eb8c8): report-1.png, report-2.png (synthetic report pages, A4 at 150 dpi); paper-1.png, paper-3.png, paper-10.png (pages 1, 3 and 10 of Unlimited-OCR-2.pdf from the release at 0480966938291445591f69f08a13b041c073a96a, 150 dpi, PyMuPDF 1.27.2.2).

Stage dumps (tools/zig/uocr_reference.py dump MODEL OUT PAGE[,PAGE...], torch 2.15.0 on the Apple CPU; the script is unchanged between c7cd2ff4a and b68ea38f2): <case>.safetensors + <case>.json (bf16) and <case>-f32.safetensors + <case>-f32.json (float32), for report-1, report-2, paper-1, paper-3, paper-10, and document (the five pages in that order as one request). Each .json holds the prompt ids, the greedy reply ids, its text and top2: each step's two highest logits as [id, logit, id, logit].

Block cases (uocr_reference.py blocks MODEL report-1.safetensors OUT --call C --view V): report-1-blocks-tile1.safetensors (call 0, view 1: tile view 1, 640) and report-1-blocks-global.safetensors (call 1, view 0: the global view, 1024).

Commands (MODEL = the release's checkpoint directory, P = the pages):

tf-uocr-decode MODEL P/report-1.png report-1.safetensors report-1-f32.safetensors
tf-uocr-decode MODEL P/report-1.png report-1.safetensors report-1-f32.safetensors --their-embeds
# likewise report-2, paper-1, paper-3, paper-10; the document adds --reply-reported to the first form:
tf-uocr-decode MODEL P/report-1.png,P/report-2.png,P/paper-1.png,P/paper-3.png,P/paper-10.png document.safetensors document-f32.safetensors --reply-reported
tf-uocr-blocks MODEL report-1-blocks-tile1.safetensors
tf-uocr-blocks MODEL report-1-blocks-global.safetensors

Unlimited-OCR-2: the OCR cases' inputs (#2074 at 0af3eb8c8)

Choose a tag to compare

@BobClawblaw BobClawblaw released this 05 Nov 22:58

Inputs for Unlimited-OCR-2's six OCR cases (ashhart#2074, head 0af3eb8c8): the five page images and the six reference transcriptions that the recipe harness's ocr plan reads. These are the exact files behind every M5 Ultra ocr run quoted on the PR.

Pages (from the recipe's ocr/pages.py):

  • report-1.png, report-2.png: synthetic report pages, A4 at 150 dpi.
  • paper-1.png, paper-3.png, paper-10.png: pages 1, 3 and 10 of Unlimited-OCR-2.pdf, the paper shipped in the release at 0480966938291445591f69f08a13b041c073a96a, rendered at 150 dpi with PyMuPDF 1.27.2.2.

References (from the recipe's ocr/refs.py): the release's own infer (modeling_unlimitedocr.py) in torch 2.15.0 bf16 on an Apple arm64 CPU, transformers 4.52.1, greedy, crop_mode=True (gundam, the publisher's server default). Each <case>.txt is the reply for <case>.png. document.txt is the five pages as one request, in the order report-1, report-2, paper-1, paper-3, paper-10.

The same release code under torch on x86 CPUs gives different documents, so compare against these files rather than regenerating the references on another machine.

file bytes sha256
report-1.png 60967 4bf458382a0a03004eca70987b504b09244bbb84a6b368cd1dac422268d96d00
report-2.png 51244 e0c165c2ea90f9f511fd3ed8dd0d1f00d39bbbdad2f41294b108f43f427e753f
paper-1.png 516919 b892374f079051cee7fc4344643fbd2d2e5772d42595a2565b4d3ebe3da311bb
paper-3.png 432135 c5867a9706c5649c37565905f705c279bb870aeedff21c00f404dfef32f31aa7
paper-10.png 356719 052d2b3578b7394787ceaabdbb649f3ee9ecb8691e853aded2f6f86e63f69086
report-1.txt 683 58c9ad432da973e4c062bde428363be0d5800e603b26fb3750ad477273beccf8
report-2.txt 502 366f37903b4eec023fb5c9a9762f88512ce5cc621b8f637546995dc6f925bdb3
paper-1.txt 2231 76c607079fe5d7b3a0beeb61785b92e203908291ae548bb77d0b57f7c3aecac8
paper-3.txt 4310 216de8962d746e0fe3003d6c724cfefc5d01392b2633abe0e3f8fb38a8e3eaf3
paper-10.txt 6197 8d377476d6ea03c6c95d7700ed8f74ce7ff62917f45cbdc4e5a6ede2a2600d82
document.txt 13998 34699873f08879229f6e66699804e963d34279734bd6eff829e064e951dc61c8

SHA256SUMS holds the same hashes for shasum -a 256 -c.

Nemotron 3.5 Transcribe reference clips (test data for ashhart/TensorFold#1632)

Choose a tag to compare

The six 16 kHz mono WAVs the served checks of ashhart#1632 use, with SHA256SUMS, a rebuild script from the LibriSpeech FLACs, and the request settings. Audio from LibriSpeech (https://www.openslr.org/12), CC BY 4.0.

zip SHA-256: 2e4960989e18efae06941b7ed0d8fa3c57a49057ff571f09b6285dd134b9476a