# Ayb 0.5 for Armenian OCR Cleanup: How to Choose a Model Size

Compare Metric AI's Armenian models on OCR mistakes, name spelling and exact field preservation before replacing your cleanup step.

By Paras Madan | 2026-11-22


An OCR cleanup model can fix an Armenian word and quietly change the invoice number next to it. The resulting text reads better, but the record is worse. That is the failure I'd look for first when evaluating Ayb 0.5, Metric AI's Armenian model family.

[Metric AI published Ayb 0.5 on November 20](https://huggingface.co/blog/Metric-AI/ayb-05-intro). There are three sizes, and its results on Armenian benchmarks give you a reason to try them. They don't tell you which size will make the fewest costly edits on your scans.

## Start with OCR output, not raw images

For the first comparison, keep your existing OCR engine and feed its text to each Ayb model. Use the same pages, prompt and output format. If you also want to try image-to-text later, run that as a separate experiment. Otherwise, a bad field can be hard to attribute to the scan, the text recognition or the cleanup step.

Build a small set of synthetic or approved Armenian documents with manually transcribed references. Include clean labels, uneven scans, mixed Armenian and English text, and rows with hard-to-read identifiers. Clean pages matter: they reveal whether the model changes text that was already correct.

## Give the model a mistake it should leave alone

Here's an invented OCR sample. It contains a deliberately dropped Armenian letter and an identifier with a zero followed by a capital O.

```text
OCR line: Հաշվ համար: AB-0O17
Reference label: Հաշվի համար
Reference identifier: AB-0O17
```

Ask the model to repair obvious spelling errors in labels, preserve identifiers and numbers exactly, and flag anything uncertain rather than guessing. The expected label repair is `Հաշվի` in place of `Հաշվ`. The `0O` sequence must remain unchanged. Do not give the model the reference line. Keep it with the grader.

Add a second case where the OCR output has the wrong identifier but no textual evidence lets the model reconstruct it. Mark that case for review, not automatic repair. The goal is a cleanup step that knows when to stop, not one that makes every line look complete.

## Compare sizes on edits, not just benchmark rank

The [0.8B card](https://huggingface.co/Metric-AI/Ayb-0.5-0.8B) lists about 1.75 GB of BF16 weights, including the vision encoder. The [4B card](https://huggingface.co/Metric-AI/Ayb-0.5-4B) lists about 9.3 GB. Those figures describe weights, not total process memory. Measure peak memory and time per page for the setup you intend to use.

Metric AI reports higher ArmBench scores for its larger models, but notes that factual knowledge, complex reasoning and instruction following remain areas for improvement. Start with 0.8B, then run the same examples on 4B and [9B](https://huggingface.co/Metric-AI/Ayb-0.5-9B). Count repaired label errors, changes to intact labels, altered protected fields and cases correctly flagged for review. A better overall language score doesn't compensate for more corrupted identifiers.

Ask an Armenian reader to review the reference labels and any disputed repairs. Keep the protected-field check mechanical: compare the identifier character for character before accepting the line. A human should not need to catch a zero-to-letter substitution by eye.

If the small model handles simple corrections, don't move every page to a larger one. Consider routing only the layout or language patterns that consistently improve on the larger model. Keep a human check when the source is unclear.

## Keep speech and licensing separate

Metric AI also published [OhanASR](https://huggingface.co/blog/Metric-AI/ohan-asr-v01), a separate Armenian speech recognition model. If your input is audio rather than a scanned document, measure transcription errors before measuring text cleanup. An improvement after editing doesn't establish that Ayb produced the transcript.

The three Ayb weight cards list CC BY-NC 4.0. Downloadable weights aren't blanket permission for commercial use, even though the underlying Qwen base has a different license. Review the terms before planning a business deployment.

For a first evaluation, I'd save the original OCR line beside the proposed correction and the review decision. Choose the smallest model that preserves the fields your application depends on, then keep measuring those errors when you change the model or the scans.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
