Skip to content

Document parsing pipeline 0.1.0-rc.3

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 03 Nov 00:59
· 22 commits to main since this release

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.