Qari-OCR CrispEmbed GGUF

GGUF conversion of NAMAA-Space/Qari-OCR-0.2.2.1-VL-2B-Instruct (Apache-2.0) for use with CrispEmbed.

Model

Arabic OCR with full diacritics (tashkeel) support. Fine-tuned from Qwen2-VL-2B-Instruct via LoRA (r=16, alpha=16, 324 adapter pairs) on 50K Arabic OCR samples.

Architecture: Qwen2-VL-2B (32L ViT 1280d + spatial merger + 28L Qwen2 LLM 1536d, GQA 12/2)
Parameters: 2B
Performance: WER=0.221, CER=0.059, BLEU=0.597
Training: 50K Arabic OCR records, 1 epoch, LoRA on attention+MLP

Files

File	Type	Size
qari-ocr-2b-f16.gguf	F16	4.7 GB
qari-ocr-2b-q8_0.gguf	Q8_0	2.3 GB
qari-ocr-2b-q4_k.gguf	Q4_K	1.6 GB

Usage

Uses the same qwen2vl engine as other Qwen2-VL models in CrispEmbed.

Conversion

LoRA adapter merged into full-precision Qwen2-VL-2B-Instruct base weights (324 pairs, tensor-by-tensor), then converted to GGUF via CrispEmbed converter. Quantized with crispembed-quantize (vision weights at Q8_0 floor).

License

Apache-2.0 (NAMAA-Space/Qari-OCR).

Provenance and EU AI Act Art. 53 note

Upstream model: NAMAA-Space/Qari-OCR-0.2.2.1-VL-2B-Instruct — published by NAMAA-Space.
Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.