---
library_name: pytorch
pipeline_tag: translation
tags:
- arabic
- english
- machine-translation
---

# Pixel Translation checkpoints

Final checkpoints for the Base, Vocab+, Byte, and Contextual Pixel Arabic-to-English
translation models. These weights are intended for use with
[mbzuai-nlp/arapixel-mt](https://github.com/mbzuai-nlp/arapixel-mt).

## Layout

Each model has MSA-pretrained checkpoints for 5%, 10%, 25%, 50%, and the full
MSA training set:

```text
<model>/msa/{5pct,10pct,25pct,50pct,full}/pytorch_model.bin
```

The curriculum checkpoints were fine-tuned on the 1:1 MSA/dialect mixture:

```text
<model>/finetuned/{5pct,10pct,25pct,50pct}/msa_dialect_1x/best_model.bin
```

Full-data checkpoints have four fine-tuning variants:

```text
<model>/finetuned/full/dialect/best_model.bin
<model>/finetuned/full/msa_dialect_1x/best_model.bin
<model>/finetuned/full/msa_dialect_2x/best_model.bin
<model>/finetuned/full/msa_dialect_4x/best_model.bin
```

`<model>` is one of `pixel`, `bytes`, `vocab`, or `base`. In the codebase,
the Base model is selected with `--model decoder`.

Every checkpoint directory also contains `training_complete.json`, which lets
evaluation select the fixed MSA or dynamic fine-tuning source layout. The shared
Vocab+ tokenizer is stored in `vocab/tokenizer/`.

## Download one checkpoint

```bash
hf download shaheen6/PixelTranslationCheckpoints   --include "pixel/msa/full/*"   --local-dir checkpoints
```

The downloaded model path is then:

```text
checkpoints/pixel/msa/full/pytorch_model.bin
```

Pass that file to `--model-checkpoint`. For Vocab+, also download the tokenizer:

```bash
hf download shaheen6/PixelTranslationCheckpoints   --include "vocab/tokenizer/*"   --local-dir checkpoints

export VOCAB_TOKENIZER_DIR="$PWD/checkpoints/vocab/tokenizer"
```