💬 NLP

GRDD+: An Extended Greek Dialectal Dataset with Cross-Architecture Fine-tuning Evaluation

This paper introduces GRDD+, a significantly expanded Greek dialectal dataset encompassing 10 varieties and over 6.3 million words, and evaluates its impact through fine-tuning experiments on three LLM architectures compared against frontier models.

Original authors: Stergios Chatzikyriakidis, Dimitris Papadakis, Sevasti-Ioanna Papaioannou, Erofili Psaltaki

Published 2026-03-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stergios Chatzikyriakidis, Dimitris Papadakis, Sevasti-Ioanna Papaioannou, Erofili Psaltaki

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of Artificial Intelligence (AI) as a massive, super-smart library. For a long time, this library was stocked almost entirely with books written in "Standard Modern Greek"—the formal, textbook version of the language used in schools, news, and official documents.

But Greece is a country of many voices. Just like a family where everyone has their own unique accent, slang, and storytelling style, Greek has dozens of dialects. Some are spoken on islands, some in mountain villages, some in Italy, and some are even ancient languages that have survived for thousands of years.

The problem? The AI library was missing almost all of these unique voices. If you asked the AI to tell a story in a specific local dialect, it would sound robotic, confused, or just plain wrong.

This paper, titled GRDD+, is like a massive renovation project for that library. Here's the simple breakdown of what the researchers did:

1. The "Great Collection" (The Dataset)

The researchers went out and gathered a huge amount of text from the internet, old books, songs, and folktales. They didn't just stick to the main four dialects they had before; they went hunting for six new, rare varieties of Greek, including:

Greco-Corsican: A Greek dialect spoken by a tiny community in Corsica (France) that is now extinct.
Griko: A Greek dialect spoken in Southern Italy.
Tsakonian: A very old, unique dialect that is so different it's almost a separate language.
Katharevusa: A "purist" version of Greek used in official writing for a long time.

They ended up with a massive collection of 6.3 million words covering 10 different varieties. Think of this as filling the library with thousands of new books written in the authentic voices of real people, not just textbooks.

2. The "Training Camp" (Fine-Tuning)

Having the books is great, but the AI (the "student") needs to learn how to read them. The researchers took three popular AI models (think of them as three different students: Llama-3, Llama-3.1, and Krikri) and put them through a special training camp.

They fed these models the new dialectal data. It's like taking a student who only knows how to speak formal English and teaching them to speak with a heavy Scottish broom, a Southern US drawl, and a Cockney accent all at once.

3. The "Taste Test" (Evaluation)

Once the training was done, the researchers put the AI to the test. They asked the AI to write short stories, dialogues, and creative pieces in these dialects.

Then, they brought in native speakers (real humans who grew up speaking these dialects) to act as judges. The judges gave the AI a score from 1 to 5:

1: "This sounds completely fake and robotic."
5: "This sounds exactly like a native speaker; I can't tell it's a machine."

4. The Surprising Results

Here is where the story gets interesting:

The "Specialist" Student Failed: One of the students, Krikri, was built specifically for Greek. You'd think it would be the best. But surprisingly, it didn't perform the best in the dialect tests. It turns out, just knowing the "main" language well doesn't automatically make you good at the local accents.
The "Generalist" Students Shined: The other two models (Llama), which weren't built specifically for Greek, actually learned the dialects better after training.
The "Big Boss" Models: The researchers also compared their trained models against the world's most powerful, expensive AI models (like the latest versions of ChatGPT, Claude, and Gemini).
- The Good News: Their smaller, trained models often beat the big, expensive ones in specific dialects! It proves that specialized training is more important than just having a giant brain.
- The Bad News: Some of the newest "Big Boss" models (like the latest Gemini) actually got worse at handling dialects, likely because they were trained to be more "standard" and lost their ability to handle local variations.

5. The "Magic of Small Data"

One of the coolest findings was about Northern Greek. The researchers only had a tiny amount of data for this dialect (about 333 examples). You'd think the AI would fail miserably. Instead, it performed almost as well as it did with the massive datasets!

The Analogy: It's like teaching someone to drive. You don't need to drive 10,000 miles to learn the basics; sometimes, a few hours of focused, high-quality practice in the right conditions is enough to get you driving safely.

Why Does This Matter?

This paper is a big deal because it shows that AI doesn't have to be one-size-fits-all. By giving AI access to the rich, messy, beautiful diversity of human language, we can make it more useful for everyone, not just the people who speak the "standard" version.

It's a step toward an AI that doesn't just sound like a robot reading a dictionary, but one that can sit down with a grandparent in a village in Crete or a fisherman in Cyprus and actually understand them.

1. Problem Statement

Modern Greek exhibits significant dialectal variation across geographical regions (e.g., Cretan, Cypriot, Pontic, Northern Greek). However, computational resources for these dialects are scarce, limiting the study and processing of regional linguistic varieties. While Large Language Models (LLMs) have advanced rapidly, they struggle with low-resource dialects, often failing to generate contextually appropriate responses or accurately identify dialects compared to Standard Modern Greek (SMG). Existing datasets are either too small, lack diversity, or do not support the fine-tuning of modern LLMs. There is a critical need for a large-scale, diverse dataset to evaluate and improve model performance on Greek dialects.

2. Methodology

A. Dataset Construction (GRDD+)

The authors extended the existing GRDD corpus to create GRDD+, resulting in a dataset of 6,374,939 words covering 10 varieties.

Expansion: The original four dialects (Cretan, Pontic, Northern, Cypriot) were enriched.
New Varieties: Six new varieties were added:
1. Greco-Corsican: An extinct variety from Corsica.
2. Griko: A minority language in Southern Italy.
3. Heptanesian: From the Ionian Islands (Venetian influence).
4. Tsakonian: A highly divergent dialect derived from Ancient Doric.
5. Maniot: From the Laconian Mani region.
6. Katharevusa: A "purist" historical variety used as the official written language until 1976.
Data Sources: Web-scraped texts (blogs, literature, songs, poems) and OCR-processed books.
Preprocessing: Removal of metadata, URLs, special characters, and duplicates.
Fine-tuning Subset: For the experiments, a subset of 20,116 examples was constructed from four dialects (Cretan, Pontic, Northern, Cypriot) using a sliding window approach (100-word chunks) to create prompt-completion pairs.

B. Model Architecture and Training

Base Models: Three 8B parameter models were selected:
1. Llama-3-8B: Multilingual instruction-tuned.
2. Llama-3.1-8B: Extended context (128k tokens).
3. Krikri-8B: A Greek-specialized model built on Llama-3.1, trained on 56.7B Greek tokens.
Fine-tuning Technique: Low-Rank Adaptation (LoRA) was used for efficiency.
- Configuration: Rank ( $r$ ) = 16, Alpha ( $\alpha$ ) = 32, Dropout = 0.1.
- Target Modules: All projection layers ($q, k, v, o, gate, up, down$).
- Trainable Parameters: ~0.8% of the base model.
Training Setup: 3 epochs, batch size 16 (effective), learning rate $3e^{-4}$ , using NVIDIA A100 GPUs.

C. Evaluation

Comparators: The fine-tuned models were compared against their base versions and three frontier models: Claude-3.7-Sonnet, Gemini-2.5, and ChatGPT-5.
Task: Generation of texts based on 7 different prompt types (short/medium/long stories, dialogues, creative writing) per dialect.
Metrics: Native speakers evaluated generated texts on a 5-point Likert scale (1 = Not dialectal, 5 = Native-level).
Reliability: Inter-rater reliability was measured using Krippendorff's Alpha (0.37–0.55), ICC(3,1) (0.87–0.96), and Weighted Cohen's Kappa.

3. Key Contributions

GRDD+ Dataset: The creation of the largest and most diverse Greek dialectal dataset to date, doubling the size of the previous corpus and adding six previously unrepresented varieties.
Cross-Architecture Benchmark: A comprehensive evaluation of how different model architectures (multilingual vs. monolingual-specialized) respond to dialectal fine-tuning.
Empirical Evidence on Data Efficiency: Demonstrating that high-quality dialectal fine-tuning can yield significant performance gains even with relatively small datasets (e.g., Northern Greek with only 333 examples).
Insight into Linguistic Distance: Preliminary analysis suggesting a correlation between model performance and the linguistic distance of the dialect from Standard Modern Greek.

4. Results

A. Impact of Fine-tuning

Significant Improvement: All fine-tuned models outperformed their base versions by approximately 1.5 to 2.0 points on the 5-point scale.
Base Model Failure: Base models (including the Greek-specialized Krikri) showed near-zero dialectal capability (scores ~1.0–1.5), highlighting that pre-training on general Greek text does not inherently capture dialectal nuances.

B. Model Performance Comparison

Frontier Models:
- Claude-3.7-Sonnet: Consistently high performance, topping Cretan (3.79) and Northern (3.86).
- ChatGPT-5: Decent but inconsistent; strong on Cypriot/Northern, weaker on others.
- Gemini-2.5: Consistently mediocre to poor performance across all dialects.
Fine-tuned Models:
- Llama-3.1-8B (Fine-tuned): Achieved the highest scores for Cypriot (3.51) and Pontic (2.86), outperforming frontier models on these specific dialects.
- Krikri-8B (Fine-tuned): Surprisingly underperformed compared to the multilingual Llama models, performing best only on Northern Greek. This suggests that explicit pre-training on Modern Greek does not guarantee superior dialectal adaptation compared to multilingual models.

C. Data Size vs. Performance Anomalies

Northern Greek: Despite having the smallest dataset (333 examples, 1.7% of the training set), it achieved strong scores (2.84–3.86), comparable to Cretan (which had 12x more data).
Pontic: Had a moderate dataset size (4,190 examples) but consistently scored the lowest (2.14–2.86).
Interpretation: The authors hypothesize that linguistic distance from Standard Modern Greek and data quality (genre diversity) play a larger role than raw token count. Northern Greek is linguistically closer to SMG, while Pontic is more divergent.

5. Significance

NLP Advancement: This work proves that targeted fine-tuning with high-quality, even small-scale, dialectal data can enable smaller open-source models to rival or exceed closed frontier models on specific low-resource tasks.
Resource Preservation: The inclusion of endangered and extinct varieties (Greco-Corsican, Tsakonian) provides a crucial digital archive for linguistic preservation.
Future Research: The dataset and findings establish a foundation for further research in sociolinguistics, code-switching detection, and the development of multi-dialectal models. The paper highlights the need for better automatic evaluation metrics and the investigation of why specialized models (Krikri) sometimes underperform multilingual ones in dialectal tasks.

In conclusion, GRDD+ is a pivotal resource that bridges the gap between low-resource dialects and modern LLM capabilities, demonstrating that with the right data and fine-tuning strategies, models can effectively learn and generate diverse Greek dialects.