Nawah-VL-25M

An Arabic image-captioning VLM, assembled from existing parts rather than trained from scratch.

Part	Source	Params	Trained
Vision tower	SigLIP2 NaFlex from LiquidAI/LFM2.5-VL-450M, patch 16	85.8M	frozen
Connector	Lfm2VlMultiModalProjector, 3072→2048→384	7.1M	yes
Language model	oddadmix/Emhotob-25M-v2 (Llama, hidden 384, 8 layers, 2048 ctx)	25.3M	stage 2

118M total. Generates Modern Standard Arabic captions for a single image.

Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image

model = AutoModelForImageTextToText.from_pretrained("oddadmix/Nawah-VL-25M")
proc  = AutoProcessor.from_pretrained("oddadmix/Nawah-VL-25M")

inputs = proc(text=["<s><image>"], images=[[Image.open("photo.jpg").convert("RGB")]],
              return_tensors="pt", padding=True, padding_side="left")
out = model.generate(**inputs, max_new_tokens=40)
print(proc.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

There is no chat template — the prompt is literally BOS followed by one <image> token, which the processor expands into 32–64 placeholders depending on aspect ratio.

Results

Held-out test split of visheratin/laion-coco-nllb:

metric	value
chrF++	27.4
BLEU	8.2
perplexity	12.2
grounding gap	2.09
distinct captions	98.7%

Grounding gap is eval loss with images rotated by one position minus loss with correct pairings. Near zero would mean the model ignores the image and generates from a language prior; 2.09 says the captions are driven by pixels.

Training

Two stages on 405,526 image/caption pairs, from oddadmix/laion-coco-nllb-arabic-filtered (the filtered arb_Arab slice of LAION-COCO-NLLB; 753,883 rows survive filtering, of which 405,526 images were still reachable):

1. Align — projector only, 2000 updates, LR 1e-3
2. Finetune — projector + language model, 8000 updates, LR 2e-4 / 5e-5

The projector's linear_1 is warm-started from LFM2.5-VL (identical 2048×3072 shape); linear_2 is randomly initialised and then rescaled so visual tokens enter the residual stream at the same magnitude as text tokens. Without that calibration the visual tokens were ~36× larger and alignment stalled.

Limitations

25.3M parameters in the language model (~13M non-embedding) is far below any standard VLM backbone. It names the main subject and its colour reliably and stops cleanly, but reads text in images poorly, sometimes invents a plausible brand, and its Arabic agreement occasionally slips. The training captions are themselves doubly machine-generated (BLIP English → NLLB Arabic), which caps how much reference-based metrics mean.

The 50M variant

oddadmix/Nawah-VL-50M is the same recipe on a larger Arabic backbone and is better on every measure taken. It is released.

metric	this model (25M)	50M
chrF++	27.4	29.0
BLEU	8.2	9.5
perplexity	12.2	8.6
grounding gap	2.09	2.36

Reproducing

Everything needed is in training/ in this repo, and the filtered corpus is published separately. training/HOWTO.md covers the four non-obvious parts.

A 50M-backbone version is available at oddadmix/Nawah-VL-50M.

Attribution

The language model is oddadmix/Emhotob-25M-v2, trained from scratch on Arabic. The vision tower is SigLIP2 as distributed in LiquidAI/LFM2.5-VL-450M, used unmodified and frozen; those weights remain LiquidAI's and their licence terms apply to that component.

License

cc-by-nc-4.0, inherited from the training data — non-commercial use only.