---
title: "Averroes at ImageEval 2025 Shared Task: Advancing Arabic"
description: "The Averroes team entered the ImageEval 2025 shared task on Arabic image captioning. 19 references, cited by 3 papers here."
authors: ["Mariam Saeed", "Sarah Elshabrawy", "Abdelrahman Hagrass", "Mazen Yasser", "Ayman Khalafallah"]
year: 2025
venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks"
source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.59/"
doi: "10.18653/v1/2025.arabicnlp-sharedtasks.59"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/saeed-2025-averroes-imageeval-2025-shared-task/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/saeed-2025-averroes-imageeval-2025-shared-task/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/12/) / Averroes at ImageEval 2025 Shared Task: Advancing Arabic Image Captioning with Augmentation and Two-Stage Generation

# Averroes at ImageEval 2025 Shared Task: Advancing Arabic Image Captioning with Augmentation and Two-Stage Generation

**Authors:** Mariam Saeed, Sarah Elshabrawy, Abdelrahman Hagrass, Mazen Yasser and [Ayman Khalafallah](https://md.research.elamur.ai/free/en/authors/ayman-khalafallah/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.59/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.59.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.59](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.59) · [OpenAlex](https://openalex.org/W4416037782)

**Citations:** 0 citations (OpenAlex, 2026-10-14) · [3](https://md.research.elamur.ai/free/en/papers/saeed-2025-averroes-imageeval-2025-shared-task/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The Averroes team entered the ImageEval 2025 shared task on Arabic image captioning. Arabic is a hard setting for this because labelled data is scarce and captions need cultural sensitivity. The team fine-tuned Qwen2.5-VL-7B with LoRA and improved it in stages.

- **Smaller model:** they tried BLIP and dropped it because it missed fine detail.
- **Augmentation:** they compared plain image transforms with a filtered scheme. In the filtered scheme, Aya-Vision-8B rewrites each caption, and a rewrite survives only if it stays close to the original by BLEU.
- **Two-model pipeline:** the first model writes a long Arabic description of the picture. The second model receives the image plus that description and outputs the short caption.
- **Cleanup:** regex rules remove repeated phrases.

The final system took first place on LLM-as-a-Judge and in the organizers' manual evaluation, and second place on Cosine Similarity. The result suggests that an explicit intermediate description, plus carefully filtered synthetic data, helps a mid-sized VLM in a low-resource language.

## Key points
- **Data:** 3,471 manually captioned images (2,718 train, 753 test). The team held out 10% of the training portion for validation.
- **Aug1:** three random Albumentations transforms per image, with captions left unchanged. It slightly lowered the scores.
- **Aug2:** three Aya-Vision-8B caption variants per image, kept only at BLEU ≥ 0.75, each paired with one image transform. This added 814 samples.
- **Results (BLEU / Cosine / LLM-as-a-Judge):**

  | System | BLEU | Cosine | LLM-as-a-Judge |
  |---|---|---|---|
  | Organizers' full fine-tune | 16.89 | 58.46 | 30.82 |
  | LoRA baseline | 22.84 | 56.95 | 30.19 |
  | Final system | 24.39 | 58.55 | 33.97 |
  | Merged two-model variants (MergeKit, TIES) | 24.55 | 58.53 | 33.11 |

  Merging did not improve on the final system.
- **Training setup:** LoRA rank 8 on all modules, effective batch size 16, learning rate 2e-5 with a cosine schedule, bf16, one A100 80GB GPU. The checkpoint with the lowest validation loss was submitted.
- **Manual evaluation:** the team led on Cultural Relevance (3.63) and Conciseness (3.43). On Completeness (2.60) and Accuracy (2.80) it trailed VLCAP (2.67 and 2.97). VLCAP also led on Cosine Similarity (60.01).

## Interesting ideas
- Splitting generation into a verbose "describe everything" step and a separate "condense into the target form" step gave the best semantic scores, because the second model works from an explicit, inspectable intermediate.
- Synthetic training variants were accepted only if they stayed close to the gold reference, which trades diversity for safety against drift.
- The intermediate descriptions were produced by a teacher that could see the target caption, so each rationale was steered toward the right answer.
- Perturbing only the input while keeping the target fixed (Aug1) gave no reliable gain, so adding more data was not automatically helpful.
- A cheap deterministic pass removing repeated phrases fixed a common decoder artifact without retraining anything.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (back): generation from the world graph could first verbalise every relevant fact in full, then hand that to a separate compression step that picks what the short answer or summary keeps, making both stages explainable.
- **I2** (learn): LLM-teacher paraphrases of treebank sentences could be kept only when their frames match the original's meaning graph, giving silver data with a built-in sanity filter.
- **I3** (answer): when building training traces for readers of fables or word problems, let the teacher see the gold answer while writing the reasoning path, then train the student to reproduce that path without it.
- **I4** (learn): every augmentation scheme should be checked against held-out scores and a no-augmentation control before it is adopted.
- **I5** (back): generated text could pass through a rule-based check for duplicated clauses or mentions before output, with each removal logged for explainability.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works

🧭 Topics (2): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works · [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 657 works

## Cite

Cite as: Mariam Saeed, Sarah Elshabrawy, Abdelrahman Hagrass, Mazen Yasser and Ayman Khalafallah. 2025. Averroes at ImageEval 2025 Shared Task: Advancing Arabic Image Captioning with Augmentation and Two-Stage Generation. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{saeed2025averroes,
  title = {{Averroes at ImageEval 2025 Shared Task: Advancing Arabic Image Captioning with Augmentation and Two-Stage Generation}},
  author = {Mariam Saeed, Sarah Elshabrawy, Abdelrahman Hagrass, Mazen Yasser and Ayman Khalafallah},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-sharedtasks.59},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.59}
}
```

- [← previous Athar at QIAS2025: LLM-based Question Answering Systems for Islamic Inheritance and Classical Islamic Knowledge](https://md.research.elamur.ai/free/en/papers/noureldien-2025-athar-qias2025-llm-based-question/)
- [next → Baseer: A Vision-Language Model for Arabic Document-to-Markdown OCR](https://md.research.elamur.ai/free/en/papers/hennara-2025-baseer-vision-language-model-arabic/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/12/)

## References

19 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, and 1 others. 2022. [Flamingo: a visual language model for few-shot learning](https://md.research.elamur.ai/free/en/papers/alayrac-2022-flamingo-visual-language-model-few/). Advances in neural information processing systems, 35:23716– 23736.
2. Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077–6086. [find](https://scholar.google.com/scholar?q=Peter+Anderson%2C+Xiaodong+He%2C+Chris+Buehler%2C+Damien+Teney%2C+Mark+Johnson%2C+Stephen+Gould%2C+and+Lei+Zhang%2E+2018%2E+Bottom%2Dup+and+top%2Ddown+attention+for+image+captioning+and+visual+question+answering%2E+In+Proc)
3. Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. [Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond](https://md.research.elamur.ai/free/en/papers/bai-2023-qwen-vl-versatile-vision-language/). arXiv preprint arXiv:2308.12966. Ahlam Bashiti, Alaa Aljabari, Hadi Hamoud, Md. [↗](https://arxiv.org/abs/2308.12966)
4. Rafiul Biswas, Bilal Shalash, Mustafa Jarrar, Fadi Zaraket, George Mikros, Ehsaneddin Asgari, and Wajdi Zaghouani. 2025. [ImageEval 2025: The First Arabic Image Captioning Shared Task](https://md.research.elamur.ai/free/en/papers/bashiti-2025-imageeval-2025-first-arabic-image/). In Proceedings of the Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou, China. Association for Computational Linguistics.
5. Marcella Cornia, Matteo Stefanini, Lorenzo Baraldi, and Rita Cucchiara. 2020. Meshed-memory transformer for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10578–10587. [find](https://scholar.google.com/scholar?q=Marcella+Cornia%2C+Matteo+Stefanini%2C+Lorenzo+Baraldi%2C+and+Rita+Cucchiara%2E+2020%2E+Meshed%2Dmemory+transformer+for+image+captioning%2E+In+Proceedings+of+the+IEEE%2FCVF+conference+on+computer+vision+and+pattern+r)
6. Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vladimir Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. 2024. [Arcee’s MergeKit: A toolkit for merging large language models](https://md.research.elamur.ai/free/en/papers/goddard-2024-arcee-s-mergekit-toolkit-merging/). In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 477–485, Miami, Florida, US. Association for Computational Linguistics. [↗](https://doi.org/10.13140/rg.2.2.10481.70249)
7. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. [Lora: Low-rank adaptation of large language models](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). In ICLR.
8. Andrej Karpathy and Li Fei-Fei. 2017. Deep visual-semantic alignments for generating image descriptions. IEEE Trans. Pattern Anal. Mach. Intell., 39(4):664–676. [↗](https://doi.org/10.1109/tpami.2016.2598339)
9. Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. 2024. [Llms-as-judges: A comprehensive survey on llm-based evaluation methods](https://md.research.elamur.ai/free/en/papers/li-2024-llms-judges-comprehensive-survey-llm/). Preprint, arXiv:2412.05579. neural information processing systems, 36:34892– 34916. [↗](https://arxiv.org/abs/2412.05579)
10. Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. 2022. PEFT: State-of-the-art parameter-efficient fine-tuning methods. [https://github.com/huggingface/peft](https://github.com/huggingface/peft). [↗](https://github.com/huggingface/peft)
11. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. [Bleu: a method for automatic evaluation of machine translation](https://md.research.elamur.ai/free/en/papers/papineni-2002-bleu-method-automatic-evaluation-machine/). In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, page 311–318, USA. Association for Computational Linguistics.
12. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. [Learning transferable visual models from natural language supervision](https://md.research.elamur.ai/free/en/papers/radford-2021-learning-transferable-visual-models-natural/). In International conference on machine learning, pages 8748–8763. PmLR.
13. Gerard Salton and Christopher Buckley. 1988. Term-weighting approaches in automatic text retrieval. Inf. Process. Manage., 24(5):513–523. [↗](https://doi.org/10.1016/0306-4573(88)90021-0)
14. Karen Sparck Jones. 1988. A statistical interpretation of term specificity and its application in retrieval, page 132–142. Taylor Graham Publishing, GBR. [find](https://scholar.google.com/scholar?q=Karen+Sparck+Jones%2E+1988%2E+A+statistical+interpretation+of+term+specificity+and+its+application+in+retrieval%2C+page+132%E2%80%93142%2E+Taylor+Graham+Publishing%2C+GBR%2E)
15. Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. [Show and tell: A neural image caption generator](https://md.research.elamur.ai/free/en/papers/vinyals-2014-show-tell-neural-image-caption/). In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156–3164. [↗](https://doi.org/10.48550/arxiv.1411.4555)
16. Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. [Show, attend and tell: neural image caption generation with visual attention](https://md.research.elamur.ai/free/en/papers/xu-2015-show-attend-tell-neural-image/). In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ICML’15, page 2048–2057. JMLR.org. [↗](https://doi.org/10.48550/arxiv.1502.03044)
17. Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. [Ties-merging: Resolving interference when merging models](https://md.research.elamur.ai/free/en/papers/yadav-2023-ties-merging-resolving-interference-merging/). Preprint, arXiv:2306.01708. [↗](https://arxiv.org/abs/2306.01708)
18. Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. [Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation](https://md.research.elamur.ai/free/en/papers/li-2022-blip-bootstrapping-language-image-pre/). arXiv preprint. [↗](https://doi.org/10.48550/arxiv.2201.12086)
19. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. [Visual instruction tuning](https://md.research.elamur.ai/free/en/papers/liu-2023-visual-instruction-tuning/). Advances in [↗](https://doi.org/10.48550/arxiv.2304.08485)

*Updated 1 November 2026*