BayanSimplify-v0.3
2
3 commits
1 repos using this model
1 linked in READMEs
updated Oct 5, 2026
Model 3 of the Bayan project and our best model: AraT5v2 trained on Bayan corpus v1, decoded with four beams.
Trained on Bayan corpus v1 (14,975 rows written by Gemma 4 31B and checked by a judge model), which teaches restructuring, restraint on easy text and keeping protected text intact. Decoded with four beams, it outperforms every earlier Bayan model.
BayanBench v2.0, test, core items, int8, four beams, through the app's text step:
| BayanSimplify-v0.3 | Earlier models | |
|---|---|---|
| Meaning kept | 82% | at most 78.6% |
| Longest clause, words cut | 5.2 | BayanSimplify-v0.2 in the app: 1.9 |
| Numbers kept | 100% |
Against BayanSimplify-v0.2 behind the same checks it is no worse on meaning and cuts about three times as many words. In a blind human rating (18 test sentences, four team raters and a reader with dyslexia), its changed outputs were found easier 86% of the time (BayanSimplify-v0.2: 60%), none harder, and every team vote between the two chose BayanSimplify-v0.3.
Input: بسط: (without the shadda), then the text. Decoding: four beams, 3-gram blocking, up to 256 tokens (the
defaults in generation_config.json). The app's int8 bundle is in
BayanSimplify-ONNX (v0.3/).
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
repo = "Congi-libya/BayanSimplify-v0.3"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
text = "وقد أدى التوسع العمراني السريع الذي شهدته المدينة خلال العقدين الماضيين إلى ازدحام مروري خانق في ساعات الذروة."
x = tok("بسط: " + text, return_tensors="pt", max_length=256, truncation=True)
y = model.generate(**x, num_beams=4, no_repeat_ngram_size=3, max_length=256)
print(tok.decode(y[0], skip_special_tokens=True))
| Model | What it is | Base model | Training data | In the Bayan app |
|---|---|---|---|---|
| BayanSimplify-v0.1 | Model 1 | AraT5v2-base-1024 | SAMER, level 5 → 3 | not shipped |
| BayanSimplify-v0.2 | Model 2 | AraT5v2-base-1024 | 22,733-pair mix (SAMER, Baseet, DAASI) with strength tags | "Earlier large model" |
| BayanSimplify-v0.2-Fast | Model 2, compact | AraBART | the same mix | "Fast model", the default download |
| BayanSimplify-v0.3 | Model 3, our best | AraT5v2-base-1024 | Bayan corpus v1, 14,975 rows | "Large model", four beams |
| BayanSimplify-ONNX | The int8 bundles the app downloads | v0.2-Fast, v0.2, v0.3 |
CC BY-NC 4.0 (non-commercial), in line with the training data's terms. SAMER is used under the CAMeL Lab's permission to fine-tune and share weights for non-commercial use; its text is not redistributed here.
Bayan (بيان) simplifies Arabic text for readers with dyslexia, entirely on an Android phone: select text in any app, choose «تبسيط» (Simplify), and a simpler version appears over the page. Built by Team Cogni for the Samsung Innovation Campus AI capstone, 2026. Every number on this card comes from the team's final report and BayanBench v2.0, the benchmark built for the task (meaning scored by Gemma 4 31B, checked against human raters, AUC 0.85).
BayanSimplify-v0.3
2
3 commits
1 repos using this model
1 linked in READMEs
updated Oct 5, 2026
Model 3 of the Bayan project and our best model: AraT5v2 trained on Bayan corpus v1, decoded with four beams.
Trained on Bayan corpus v1 (14,975 rows written by Gemma 4 31B and checked by a judge model), which teaches restructuring, restraint on easy text and keeping protected text intact. Decoded with four beams, it outperforms every earlier Bayan model.
BayanBench v2.0, test, core items, int8, four beams, through the app's text step:
| BayanSimplify-v0.3 | Earlier models | |
|---|---|---|
| Meaning kept | 82% | at most 78.6% |
| Longest clause, words cut | 5.2 | BayanSimplify-v0.2 in the app: 1.9 |
| Numbers kept | 100% |
Against BayanSimplify-v0.2 behind the same checks it is no worse on meaning and cuts about three times as many words. In a blind human rating (18 test sentences, four team raters and a reader with dyslexia), its changed outputs were found easier 86% of the time (BayanSimplify-v0.2: 60%), none harder, and every team vote between the two chose BayanSimplify-v0.3.
Input: بسط: (without the shadda), then the text. Decoding: four beams, 3-gram blocking, up to 256 tokens (the
defaults in generation_config.json). The app's int8 bundle is in
BayanSimplify-ONNX (v0.3/).
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
repo = "Congi-libya/BayanSimplify-v0.3"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
text = "وقد أدى التوسع العمراني السريع الذي شهدته المدينة خلال العقدين الماضيين إلى ازدحام مروري خانق في ساعات الذروة."
x = tok("بسط: " + text, return_tensors="pt", max_length=256, truncation=True)
y = model.generate(**x, num_beams=4, no_repeat_ngram_size=3, max_length=256)
print(tok.decode(y[0], skip_special_tokens=True))
| Model | What it is | Base model | Training data | In the Bayan app |
|---|---|---|---|---|
| BayanSimplify-v0.1 | Model 1 | AraT5v2-base-1024 | SAMER, level 5 → 3 | not shipped |
| BayanSimplify-v0.2 | Model 2 | AraT5v2-base-1024 | 22,733-pair mix (SAMER, Baseet, DAASI) with strength tags | "Earlier large model" |
| BayanSimplify-v0.2-Fast | Model 2, compact | AraBART | the same mix | "Fast model", the default download |
| BayanSimplify-v0.3 | Model 3, our best | AraT5v2-base-1024 | Bayan corpus v1, 14,975 rows | "Large model", four beams |
| BayanSimplify-ONNX | The int8 bundles the app downloads | v0.2-Fast, v0.2, v0.3 |
CC BY-NC 4.0 (non-commercial), in line with the training data's terms. SAMER is used under the CAMeL Lab's permission to fine-tune and share weights for non-commercial use; its text is not redistributed here.
Bayan (بيان) simplifies Arabic text for readers with dyslexia, entirely on an Android phone: select text in any app, choose «تبسيط» (Simplify), and a simpler version appears over the page. Built by Team Cogni for the Samsung Innovation Campus AI capstone, 2026. Every number on this card comes from the team's final report and BayanBench v2.0, the benchmark built for the task (meaning scored by Gemma 4 31B, checked against human raters, AUC 0.85).