Arva OCR

1. Abstract

We are introducing Arva OCR, an advanced multilingual structural document parsing model developed entirely from scratch by SAGEA. In our commitment to solving regional constraints in artificial intelligence, Arva is a highly optimized 3-billion parameter architecture built to natively understand the complex spatial dimensions and topologies of South Asian scripts, without relying on legacy OCR pipelines or pre-existing base models.

Through targeted pre-training and supervised fine-tuning natively on localized corpora using our continuous generative framework, Arva establishes new state-of-the-art (SOTA) performance benchmarks. This capability extends beyond standard typography to encompass highly variable handwritten Nepali scripts, complex mixed-script document layouts, and granular preservation of layout topology.

2. Technical Specifications & GLUE Pipeline

Trained with the GLUE Engine

One of the critical reasons Arva achieves its performance on Devanagari is due to its foundational integration with SAGEA’s proprietary GLUE (Generative Ligature & Union Engine). Document parsing for languages with continuous headlines (shirorekha), conjunct consonants, and vertical modifiers cannot be reliably trained using standard bounding-box segmentation frameworks designed for Latin text.

The GLUE framework employs intense spatial and procedural optimizations, including Poisson boundary blending to enforce gradient continuity across character junctions. By utilizing GLUE during the model’s ingestion of training data, Arva mathematically eliminates the traditional junction seam artifacts that confuse typical vision encoders, allowing it to "read" Devanagari the way it is naturally written: as a continuous morphological flow rather than a string of disconnected blocks.

Unified Visual-Semantic Parsing

Conventional OCR systems frequently suffer from cascading errors derived from disjointed processing pipelines: they deploy one model for heuristic bounding-box detection, another for character classification, and a myriad of brittle post-processing scripts to attempt to reverse-engineer logical reading order.

Arva operates on a purely monolithic philosophy. We treat both text and complex visual components (like tables with merged cells, infographics, and mathematical block formulas) as first-class token sequences evaluated continuously by our cohesive vision-language model. By subverting traditional rigid coordinates in favor of direct autoregressive sequence generation, Arva natively exports structured, semantic representation strings. It intrinsically understands that a caption belongs to an image, or that a two-column newspaper must be read column-by-column rather than line-by-line across a page margin.

SVG Graphic Reconstruction

Moving past character recognition, Arva possesses the capability to logically parse and reconstruct visual data. Rather than blindly cropping a diagram and embedding it as an opaque `.png` payload, Arva interprets the geometrical constraints of simple charts or stamps and outputs corresponding renderable SVG code. This enables lossless scaling, recoloring, and direct programmatic querying of historically inaccessible graphical data within scanned PDFs.

3. State-of-the-Art Evaluation in Nepali

Generalist OCR models frequently collapse when tasked with the conjunct consonants, half-characters, and continuous top-bar lines native to Devanagari script. To validate Arva, we compiled a rigorous zero-shot evaluation harness comprised of physical municipal records, bank ledgers, and low-DPI historical newspaper scans from across Nepal.

Note: Lower Character Error Rate (CER) is superior. Higher Structural Accuracy is superior. Benchmarks conducted at int8 precision.

Most notably, Arva demonstrates an unprecedented 5.8% Character Error Rate on heavily degraded, organically hand-filled Nepali forms. By expanding the native VLM tokenizer vocabulary to faithfully represent the rich morphological structure of the Devanagari script, Arva outright sidesteps the sub-word segmentation and spatial failures frequently observed when imposing Western OCR parsers onto Indic topography.

4. Primary Use Cases

• Government & Civic Data Structuring: Converting decades of physical civic records, land registries, and census documents containing highly variable handwriting and erratic table structures into query-ready digital graphs.
• Financial and Medical Compliance: Executing high-precision extraction on localized financial disclosures and regional medical invoices, ensuring no loss of structural integrity across deeply nested line items.
• Scalable Cultural Archiving: Preserving aging literature and historically significant manuscripts susceptible to material degradation by capturing nuanced typography perfectly mapped to digital text and markdown annotations.

5. Limitations & Deployment Safety

Consistent with our commitment to transparency, we strictly enforce awareness regarding the present architectural limits of the model:

• Autoregressive Visual Confabulation: Like all generative vision transformers predicting output sequences, there exists a non-zero risk of "hallucinating" characters when source visual data is profoundly degraded or occluded. The model defaults to inferring the most statistically probable sequential characters rather than emitting an empty string. Evaluators must design systems that factor in human-in-the-loop validation for critical numerical data.
• Extreme Cursive Degradation: While providing SOTA accuracy for standard block handwriting, Arva's reliability diminishes semi-linearly when processing idiosyncratic or highly stylized "doctor's cursive" lacking distinctly isolated stroke definitions.