Skip to content

Releases

Four-model demo v5.5 with source-reference inspection

Choose a tag to compare

@bmaged23 bmaged23 released this 04 Nov 00:20

Side-by-side evaluation reference viewer for 53 shareable pages and 501 frozen assistant-created annotations. Scroll independently; hover to highlight; explicitly click either a source region or an annotation to jump to its opposite match. Page changes update both panes. Includes table grids, chart values, provenance, draft status and downloadable reference JSON. No independent human verification is implied; omitted pages remain local.

Preserves BF16 initial default, model ordering, selected model across PDFs and per-PDF page memory. README badges link to CI, Docker Hub and all three OCR model families.

SHA-256: 24f8ecebc422f250221278a0aa99a1d02828b63ef5668521b3f8badcecfcef9d

Document parsing pipeline 0.1.0-rc.4

Pre-release

Choose a tag to compare

@github-actions github-actions released this 03 Nov 11:09

Quantized dots GPU warmup fix · 0.1.0-rc.4

The real parsing worker now skips the native server's fixed-size dummy vision
warmup. On the tested GTX 1660 Ti 6 GB, rc.3 exhausted VRAM while reserving a
3,645.66 MiB dummy workspace; disabling that step lets the real synthetic page
complete. This keeps Q4_K_M language / Q8_0 projector weights, actual page pixel
limits and full GPU offload unchanged. It does not guarantee arbitrary pages fit.

The synthetic GPU check covers preparation, text/table extraction, canonical
exports, review HTML and server teardown, with no cloud calls. It is a runtime
regression probe, separate from OCR-quality evaluation. Earlier rc.3 failure
evidence is retained; CPU-hosted checks are not credited as GPU inference.

The default remains quantized dots for the 6 GB laptop. Original BF16 dots scored
best on selected handwritten mathematics (32/37 vs Q4 29/37), with 8.53 GiB peak
PyTorch allocation on Colab T4; this is not total VRAM or a universal ranking.
Optional visual descriptions use Gemma through Ollama Cloud. Weights and secrets
are not bundled. The public four-profile demo remains version 4.1.

Docker setup ·
Validation.

Public GPU parsing image: bmaged23/multimodal-document-parser@sha256:3cc1432bb586a40dffeb3757e1be0587094ecd9d4deb33fd5c0fbe5ab13ad7a4.

Built from e810b9f36f032ebdc0ce37c97f7d9730229ff43a after source/browser CI and image smoke tests. Includes SBOM and build provenance. Gemma executes through Ollama Cloud; its weights are not bundled.

Docker setup

Published-image GPU verification

The exact public rc.4 image passed a synthetic text/table page through preparation and parsing workers on GTX 1660 Ti, including canonical exports, review HTML and server teardown, with no cloud calls. Two earlier vision-weight allocation failures under Windows host-memory pressure are retained; the unchanged image succeeded after releasing WSL file caches. This is a runtime check, separate from corpus OCR quality.

Download the validation record.

Document parsing pipeline 0.1.0-rc.3

Pre-release

Choose a tag to compare

@github-actions github-actions released this 03 Nov 00:59

Quantized dots production pipeline · 0.1.0-rc.3

The parsing CLI/API defaults and parsing container now use dots.mocr Q4_K_M
language weights plus a Q8_0 vision projector. The image builds a checksum-pinned
CUDA 12.6 llama.cpp server and verifies separately downloaded model weights.
The worker owns its server for each GPU job and releases GPU memory afterward.
The image includes PDF/DOCX preparation, canonical JSON/Markdown/HTML and optional
Gemma descriptions through Ollama Cloud; credentials and model weights are not baked in.

The four-profile comparison covers 31 shared pages, with raw quality separated
from canonical acceptance. Original BF16 dots scored best on selected handwritten
mathematics (32/37 versus Q4 29/37), with 8.53 GiB peak PyTorch allocation on
Colab T4. It does not win every category. Quantized dots is the local production
choice for the tested 6 GB laptop. Historical Paddle results and image versions
remain preserved and explicitly labeled.

Public demo v4 adds original BF16 alongside the three retained profiles on NASA
and the author assignment. It defaults to quantized dots and retains the
historical Paddle IRENA/IMF pipeline stages. Public source allowlists and private
sample exclusions remain in effect.

Docker setup · All-PDF evaluation ·
Measured memory and model choice.

Public GPU parsing image: bmaged23/multimodal-document-parser@sha256:6ab09ab24db0795021e093b26ba826976db5b04374c6f1b74c7b194b80733a68.

Built from b62cf534bd059f0cc580fef26da198960a6d563b after source/browser CI and image smoke tests. Includes SBOM and build provenance. Gemma executes through Ollama Cloud; its weights are not bundled.

Docker setup

Subsequent GPU finding

The real worker probe on GTX 1660 Ti failed when the native vision dummy warmup
requested a 3,645.66 MiB CUDA workspace. Hosted CPU/API checks did not expose this.
0.1.0-rc.4 disables dummy warmup; weights, real input pixel budget and full GPU
offload stay unchanged. The corrected synthetic text/table worker probe passed.
Retained rc.3 validation.

Four-model demo v5.4 with evaluation reference viewer

Choose a tag to compare

@bmaged23 bmaged23 released this 03 Nov 23:53

Adds a read-only side-by-side evaluation reference viewer: 53 shareable source pages and 501 frozen assistant-created reference annotations. Inspect text, table grids, visual regions, chart values, provenance and draft status; download the allowlisted annotation subset as JSON. Omitted source pages and their transcriptions remain private. The complete evaluation still covers 70 retained pages.

The viewer is linked from the demo, evaluation report, README and Docker Hub overview. Existing BF16 initial default, model ordering, persistent model selection and per-document page memory are preserved.

SHA-256: 4585dbf07859d047981b0b877bf3e28627fe24aeedc49027deb7621909aed654

Four-model interactive demo v5.3

Pre-release

Choose a tag to compare

@bmaged23 bmaged23 released this 03 Nov 23:32

Start with original dots.mocr BF16. Model selector order: BF16, quantized dots.mocr Q4/Q8, PaddleOCR-VL-1.6, Unlimited-OCR 2. Switching PDFs preserves the selected OCR model and each PDF remembers its page. Explicit model/page links still work. No model outputs have been rerun or altered.

SHA-256: b8c0bafbdb229514a703e684c6969ad5b2ce0e9c4f7f7b1028a1ba899c1fed9b

Four-model interactive demo v5.2

Pre-release

Choose a tag to compare

@bmaged23 bmaged23 released this 03 Nov 23:29

Default to original dots.mocr BF16 on every document. Model selector order: BF16, quantized dots.mocr Q4/Q8, PaddleOCR-VL-1.6, Unlimited-OCR 2.

Document order and per-document page memory are unchanged. Explicit model/page links remain supported. Saved model outputs have not been rerun or altered.

SHA-256: 4e636a798e190d587b5e3881e766b0beec21ef53f995d5d382864d2f17217959

Four-model interactive demo v5.1

Pre-release

Choose a tag to compare

@bmaged23 bmaged23 released this 03 Nov 23:15

Interactive four-model OCR comparison with separate OCR and Gemma producer labels.

  • Handwritten mathematics, IMF complex tables, IRENA charts/tables, then NASA; original titles preserved.
  • Start each document at its first available page; remember its page when switching documents.
  • 53 shareable source pages, 16 model/document combinations; retained model failures remain visible.
  • Integrated evaluation report covers 70 retained pages and all four OCR profiles, with metric definitions and limitations.
  • Public bundle excludes full private reference transcriptions, failure traces and omitted source pages. No live inference.

SHA-256: 5dbad6009fa3d4f84fe9ad3a526a1679121e765393dfd44b84eb38c9e3232965

Document parsing pipeline 0.1.0-rc.2

Pre-release

Choose a tag to compare

@github-actions github-actions released this 02 Nov 11:16

GPU document-parsing pipeline — release candidate 2

The parsing runtime is available as a public Linux amd64 image:

docker pull bmaged23/multimodal-document-parser:0.1.0-rc.2

It includes separate preparation/parsing APIs and workers, PDF rendering, DOCX
conversion, PaddleOCR-VL-1.6 GPU inference, canonical JSON/Markdown/HTML exports,
and the Ollama Cloud client for optional chart/image descriptions. Gemma executes
remotely; provide your API key at runtime. The separate dots.mocr comparison
server, model weights, private inputs and the full RAG stack are not bundled.

The image was built and published by
GitHub Actions.
Source/browser CI, container runtime checks and anonymous registry access passed.
SPDX SBOM and SLSA build provenance accompany the image. Its source revision is
5d3c1ba9ccf00917307e0616292632b9e232f1e3.

Immutable image index:
sha256:fbd8e78741c933e04de58ab5648f55b3846cad65fc73290ba7ac8a35eb580c34.

Hosted runners do not provide an NVIDIA GPU. Actual CUDA tensor validation covers
the earlier local 0.1.0-rc.1 image; no fresh OCR corpus run or cloud call was made
for this release. Legacy .doc and encrypted documents are unsupported, and
extraction quality varies with layout, image quality and handwriting.

Docker installation ·
API usage ·
Parser evaluation ·
Live saved-output demo

The demo asset, when attached, is a curated presentation of previously saved
public-source results, with source mappings and a checksum inventory. It runs
no live inference and is not the complete local evaluation archive. GitHub's
source archives refer to the exact image source revision; current presentation
guides and the demo are linked separately above.

Saved-output demo data updates

public-demo-v3.zip replaces the current handwriting example with four
author-written mathematics pages and a separately scoped three-model evaluation.
Earlier v1/v2 demo assets remain available as historical snapshots. This static
demo/data update does not rebuild or change the container image above.

Parser comparison demo v4.1

Pre-release

Choose a tag to compare

@bmaged23 bmaged23 released this 02 Nov 22:16

Quantized dots production pipeline · 0.1.0-rc.3

The parsing CLI/API defaults and parsing container now use dots.mocr Q4_K_M
language weights plus a Q8_0 vision projector. The image builds a checksum-pinned
CUDA 12.6 llama.cpp server and verifies separately downloaded model weights.
The worker owns its server for each GPU job and releases GPU memory afterward.
The image includes PDF/DOCX preparation, canonical JSON/Markdown/HTML and optional
Gemma descriptions through Ollama Cloud; credentials and model weights are not baked in.

The four-profile comparison covers 31 shared pages, with raw quality separated
from canonical acceptance. Original BF16 dots scored best on selected handwritten
mathematics (32/37 versus Q4 29/37), with 8.53 GiB peak PyTorch allocation on
Colab T4. It does not win every category. Quantized dots is the local production
choice for the tested 6 GB laptop. Historical Paddle results and image versions
remain preserved and explicitly labeled.

Public demo v4 adds original BF16 alongside the three retained profiles on NASA
and the author assignment. It defaults to quantized dots and retains the
historical Paddle IRENA/IMF pipeline stages. Public source allowlists and private
sample exclusions remain in effect.

Docker setup · All-PDF evaluation ·
Measured memory and model choice.

Four-profile parser comparison demo v4

Pre-release

Choose a tag to compare

@bmaged23 bmaged23 released this 02 Nov 21:54

Quantized dots production pipeline · 0.1.0-rc.3

The parsing CLI/API defaults and parsing container now use dots.mocr Q4_K_M
language weights plus a Q8_0 vision projector. The image builds a checksum-pinned
CUDA 12.6 llama.cpp server and verifies separately downloaded model weights.
The worker owns its server for each GPU job and releases GPU memory afterward.
The image includes PDF/DOCX preparation, canonical JSON/Markdown/HTML and optional
Gemma descriptions through Ollama Cloud; credentials and model weights are not baked in.

The four-profile comparison covers 31 shared pages, with raw quality separated
from canonical acceptance. Original BF16 dots scored best on selected handwritten
mathematics (32/37 versus Q4 29/37), with 8.53 GiB peak PyTorch allocation on
Colab T4. It does not win every category. Quantized dots is the local production
choice for the tested 6 GB laptop. Historical Paddle results and image versions
remain preserved and explicitly labeled.

Public demo v4 adds original BF16 alongside the three retained profiles on NASA
and the author assignment. It defaults to quantized dots and retains the
historical Paddle IRENA/IMF pipeline stages. Public source allowlists and private
sample exclusions remain in effect.

Docker setup · All-PDF evaluation ·
Measured memory and model choice.