File size: 2,878 Bytes
a57a516
 
 
 
 
 
 
 
cc668af
 
 
 
a57a516
cc668af
 
 
efe60bb
cc668af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9d041d5
 
 
 
 
 
 
 
cc668af
9d041d5
97c4ed4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
---
title: README
emoji: 🌍
colorFrom: yellow
colorTo: gray
sdk: static
pinned: false
---
<div align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/63da3d7ae697e5898cb86854/4EnLA20pUWnvqppA5y5Q4.gif" alt="denoising_small_16_9" />
  <h1>Diffutron: A Masked Diffusion Language Model for Turkish Language</h1>
</div>

<p align="center">
  &nbsp&nbsp | 🤗 <a href="https://huggingface.co/collections/diffutron/diffutronlm">Models</a>&nbsp&nbsp | 
  &nbsp&nbsp 📊 <a href="https://huggingface.co/datasets/diffutron/DiffutronLM-Pretraining-Corpus">Pre-training Dataset</a>&nbsp&nbsp | 
  &nbsp&nbsp 📄 <a href="https://arxiv.org/abs/2603.20466">Paper</a>&nbsp&nbsp |
</p>

## Overview

Diffutron is a lightweight, non-autoregressive Masked Diffusion Language Model (MDLM) specifically optimized for the Turkish language. By utilizing a discrete diffusion process, Diffutron generates text through iterative refinement, allowing for bi-directional context awareness and high parameter efficiency.

## Core Features

* **Architecture:** Discrete Masked Diffusion (MDLM) using a 307M parameter encoder backbone.
* **Efficiency:** Achieves competitive performance against 2B+ parameter autoregressive models on Turkish benchmarks.
* **Adaptation:** LoRA-based (r=256) continual pre-training on a 2M sequence Turkish corpus.
* **Instruction Tuning:** Progressive strategy using LlamaTurk and InstrucTurca datasets for enhanced command following.

## Benchmarks

Diffutron achieves a significant reduction in perplexity and competitive scores across the CETVEL benchmark suite:

| Benchmark | Diffutron-1st-Stage (0.3B) | Diffutron-2nd-Stage (0.3B) | TURNA (1.1B) | Kumru (2B) | Kanarya (2B) | Llama-3.2 (3B) | Trendyol (7B) | Aya-101 (13B) |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **Belebele_TR** | 22.22 | 27.00 | 22.56 | 29.00 | 28.11 | **55.78** | 36.22 | 22.89 |
| **EXAMS_TR** | 25.95 | 27.74 | 23.66 | **30.03** | **30.03** | 26.21 | 28.50 | 22.90 |
| **IronyTR** | 50.67 | **52.00** | 48.33 | 51.00 | 50.00 | 50.17 | 50.00 | **52.17** |
| **News_Cat** | 23.20 | 32.40 | 32.80 | 26.40 | 66.80 | 64.00 | **81.20** | 20.00 |
| **MNLI_TR** | 33.29 | 32.81 | 34.94 | **36.42** | 33.40 | 34.76 | 35.19 | 27.90 |
| **STS_TR** | 17.77 | **18.78** | 14.21 | 11.75 | 12.91 | 12.91 | 15.52 | 16.97 |
| **XCOPA_TR** | 53.80 | 52.00 | 55.80 | 54.00 | **64.20** | 54.60 | 61.00 | 59.60 |
| **Average** | 32.41 | **34.68** | 33.19 | 34.09 | 40.78 | 42.63 | **43.95** | 31.78 |


## Citation

```bibtex
@misc{diffutron2026,
      title={Diffutron: A Masked Diffusion Language Model for Turkish Language}, 
      author={Şuayp Talha Kocabay and Talha Rüzgar Akkuş},
      year={2026},
      eprint={2603.20466},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.20466}, 
}
```