Nanonets-OCR-s

Nanonets-OCR-s#

Official Resources#

Role In Tabulus#

Nanonets-OCR-s is a vision-language table reconstruction candidate used by Tabulus for Step 2 reconstruction from canonical MinerU table crops. The Tabulus adapter sends the crop directly to the pinned Nanonets-OCR-s model checkpoint and expects native structured HTML table output.

Nanonets-OCR-s is an image-to-Markdown OCR VLM built for structured document conversion. The model card describes support for HTML/Markdown tables, LaTeX equations, image descriptions, signatures, watermarks, and checkbox handling. It is based on Qwen2.5-VL, while Tabulus treats nanonets/Nanonets-OCR-s as the adapter identity and uses its HTML table output directly.

The registered Tabulus adapter is:

nanonets-ocr-s

The exact model checkpoint used by Tabulus is nanonets/Nanonets-OCR-s at revision 3baad182cc87c65a1861f0c30357d3467e978172. Its underlying backbone architecture is Qwen2.5-VL, implemented through the runtime Transformers class Qwen2_5_VLForConditionalGeneration. Tabulus is not substituting a generic Qwen checkpoint for Nanonets-OCR-s.

Nanonets-OCR-s consumes canonical MinerU table crops through the shared Step 2 adapter contract. It preserves the model-produced Markdown/HTML-style response before shared parsing.

Invocation#

Run Nanonets-OCR-s through the shared reconstruction CLI:

export CUDA_VISIBLE_DEVICES=0

python -m tabulus.cli reconstruct-tables \
  --crops <canonical-crop-directory> \
  --adapter nanonets-ocr-s \
  --device gpu:0

For multiple canonical crop roots:

python -m tabulus.cli reconstruct-tables \
  --crops-folder <table-crops-root> \
  --adapter nanonets-ocr-s \
  --device gpu:0

Nanonets-OCR-s operates on the canonical table-crop handoff produced earlier in the Tabulus pipeline. It does not perform the initial PDF profiling or table-cropping step.

Settings Used By Tabulus#

The validated integration uses:

  • model checkpoint: nanonets/Nanonets-OCR-s

  • model revision: 3baad182cc87c65a1861f0c30357d3467e978172

  • backbone architecture: Qwen2.5-VL

  • Transformers model class: Qwen2_5_VLForConditionalGeneration

  • processor: AutoProcessor

  • processor setting: use_fast=False

  • model loader: AutoModelForImageTextToText

  • Transformers: 4.52.4

  • tokenizers: 0.21.4

  • FlashAttention: 2.7.3

  • PyTorch: 2.6.0+cu124

  • attention implementation: flash_attention_2

  • model dtype: bfloat16

  • system message: You are a helpful assistant.

  • max_new_tokens=15000

  • do_sample=False

  • registry capability: GPU only

The explicit use_fast=False processor setting is intentional. The first successful feasibility run used the default slow image processor, and setting use_fast=False reproduced that output bit-for-bit.

The validated environment name was tabulus-nanonets-ocr-s on an NVIDIA L40S. The adapter also imports PyTorch, torchvision, Pillow, and Accelerate at runtime. Those are validated environment and dependency details, not reconstruction-accuracy claims.

The implemented adapter path does not require Docker, vLLM, or qwen-vl-utils.

The prompt used by the adapter is:

Extract the text from the above document as if you were reading it naturally.
Return the tables in html format. Return the equations in LaTeX representation.
If there is an image in the document and image caption is not present, add a
small description of the image inside the <img></img> tag; otherwise, add the
image caption inside <img></img>. Watermarks should be wrapped in brackets. Ex:
<watermark>OFFICIAL COPY</watermark>. Page numbers should be wrapped in
brackets.
Ex: <page_number>14</page_number> or <page_number>9/22</page_number>. Prefer
using ☐ and ☑ for check boxes.

Tabulus explicitly runs deterministic generation with do_sample=False. Transformers may warn that a temperature value inherited from the model generation configuration is ignored; temperature is not an intentional Tabulus generation parameter for this adapter.

Input And Image Processing#

The adapter converts the supplied canonical crop to RGB and passes it through the model processor. Model-internal image preprocessing or resizing is allowed, but it is not external Tabulus recropping or table redetection.

The common-crop policy is deliberate: reconstruction candidates are compared on the same visual evidence rather than on candidate-specific detections or crops from the original PDF.

Native Output#

The validated table test produced native structured HTML. Tabulus preserves the raw decoded generation, creates the parser-facing clean generation by removing model special tokens, and sends the clean model output directly to:

tabulus.table_ocr.parsing:parse_table_text

There is no Nanonets-specific parser and no Nanonets-specific semantic or structural normalization pass. Existing deterministic HTML parsing rules apply unchanged.

The validated Nanonets output included rich HTML table constructs such as:

<table>
<thead>
<tbody>
<th>
<td>
rowspan
colspan
<sup>
<sub>
<br>

Tabulus does not semantically fix or clean the model-generated table, remove model-native blank rows, repair cell contents, merge continued tables, or perform reference-resolution heuristics during reconstruction.

Native Nanonets-OCR-s artifacts preserve model provenance and native evidence, including model repository and revision, backbone architecture, runtime model class, processor/model loader settings, prompt and system message, generation settings, dtype, attention implementation, execution device, dependency versions, raw generation, clean parser-facing generation, detected table counts, parser errors, and the canonical-crop input policy.

The adapter writes the standard reconstruction artifact layers:

<crop-root>/
  reconstructions/
    nanonets-ocr-s/
      native/
      parsed/
      predictions/
      batch_summary.json

Prediction CSV creation follows the shared Tabulus rule: the adapter result must be ok and exactly one structured table must parse from the crop. Native and parsed evidence is retained for empty or ambiguous results.

Validation Notes#

Integration validation covered deterministic direct model execution, the shared parser path, a real end-to-end CLI reconstruction, and the full multi-crop engineering run recorded in Current State. Focused Nanonets adapter tests reported 10 passed, and the complete Tabulus test suite reported 216 passed after integration.

These are engineering and integration validation observations. They are not gold-standard reconstruction accuracy, precision, recall, F1, runtime guarantees, or evidence that Nanonets-OCR-s is better or worse than another adapter.

Limitations#

This adapter reconstructs one canonical MinerU crop at a time. It does not independently locate or crop tables from the source PDF, semantically correct cell contents, merge continued tables, classify reference tables, extract bibliographies, match references, resolve DOI values, or write final resolved CSV files.