---

license: mit

base_model: facebook/m2m100_1.2B

tags:

- machine-translation

- ancient-greek

- modern-greek

- m2m100

- qlora

- peft

- natural language processing

- computational linguistics

language:

- grc

- el

pipeline_tag: translation

library_name: transformers

---

# M2M100-1.2B for Ancient Greek to Modern Greek (QLoRA)

[![DOI](https://img.shields.io/badge/DOI-10.63317%2F4cdk64dgm2w9-blue)](https://doi.org/10.63317/4cdk64dgm2w9)

This model is a fine-tuned version of [facebook/m2m100_1.2B](https://huggingface.co/facebook/m2m100_1.2B) for translating **Ancient Greek** to **Modern Greek**.

It was fine-tuned using **QLoRA (4-bit Quantization + LoRA)** on the sentence-level **AG-MG Parallel Corpus**.

The tokenizer has been expanded with **122 Ancient Greek characters** (Polytonic) that were missing from the original M2M100 vocabulary and are essential for handling the source text correctly.

This model was trained by Spyridon Mavromatis at the Institute for Language and Speech Processing (ILSP), "Athena" RC, and the National and Kapodistrian University of Athens (NKUA) as part of an M.Sc. thesis.

---

## Model Details

- **Base Model:** facebook/m2m100_1.2B

- **Method:** QLoRA (Rank=16, Alpha=32, 4-bit NF4)

- **Vocabulary:** Expanded with 122 Polytonic Greek characters.

- **Training Data:** ~130k sentence pairs from the AG-MG Corpus.

---

## Usage

You need to load the base model, resize the embeddings, and then load the Peft adapter. If you want to load the base model in 4-bit you need `bitsandbytes` installed.

```python

import torch

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, BitsAndBytesConfig

from peft import PeftModel

# 1. Configuration

adapter_repo = "ilsp/m2m100-1.2B-ag-mg-qlora"

base_model_id = "facebook/m2m100_1.2B"

# 2. Load Tokenizer (from adapter repo for added tokens)

tokenizer = AutoTokenizer.from_pretrained(adapter_repo, src_lang="el") # We treat AG as 'el' script-wise

# 3. Load Base Model in 4-bit

bnb_config = BitsAndBytesConfig(

    load_in_4bit=True,

    bnb_4bit_quant_type="nf4",

    bnb_4bit_compute_dtype=torch.bfloat16,

    bnb_4bit_use_double_quant=True

)

model = AutoModelForSeq2SeqLM.from_pretrained(

    base_model_id, 

    quantization_config=bnb_config, 

    device_map="auto"

)

# 4. Resize Embeddings (Critical)

model.resize_token_embeddings(len(tokenizer))

# 5. Load Adapter

model = PeftModel.from_pretrained(model, adapter_repo)

model.eval()

# 6. Inference

text = "Ὦ ξεῖν', ἀγγέλλειν Λακεδαιμονίοις ὅτι τῇδε κείμεθα."

inputs = tokenizer(text, return_tensors="pt").to(model.device)

# Force target language to Modern Greek ('el')

forced_bos_token_id = tokenizer.get_lang_id("el")

translated_tokens = model.generate(

    **inputs, 

    forced_bos_token_id=forced_bos_token_id, 

    max_length=128

)

print(tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0])
```

---

## Performance

### Main Test Set Results

Evaluated on the 2,000 sentence-pairs Test Set (Attic & Koine Hellenistic dialects).

| **Model** | **Method** | **BLEU ↑** | **chrF++ ↑** | **TER ↓** | **BERTScore F1 ↑** | **COMET ↑** | **ΔBLEU** |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **NLLB-600M** | [Base](https://huggingface.co/facebook/nllb-200-distilled-600M) | 1.55 | 16.86 | 106.80 | 0.880 | 0.539 | - |
|  | [LoRA](https://huggingface.co/ilsp/nllb-200-600M-ag-mg-lora) | 7.43 | 29.31 | 88.32 | 0.903 | 0.667 | +5.88 |
| **NLLB-1.3B** | [Base](https://huggingface.co/facebook/nllb-200-distilled-1.3B) | 2.15 | 17.78 | 106.41 | 0.885 | 0.573 | - |
|  | [LoRA](https://huggingface.co/ilsp/nllb-200-1.3B-ag-mg-lora) | 8.01 | 30.02 | 87.74 | 0.905 | 0.687 | +5.86 |
| **M2M100-1.2B** | [Base](https://huggingface.co/facebook/m2m100_1.2B) | 0.62 | 10.70 | 100.50 | 0.858 | 0.475 | - |
| 👉 | [QLoRA](https://huggingface.co/ilsp/m2m100-1.2B-ag-mg-qlora) | 10.96 | 33.09 | **82.99** | **0.911** | 0.710 | **+10.34** |
|  | [Full FT](https://huggingface.co/ilsp/m2m100-1.2B-ag-mg-full-ft) | 9.60 | 31.16 | 83.43 | 0.908 | 0.692 | +8.98 |
| **Krikri-8B-Instruct** | [Base](https://huggingface.co/ilsp/Llama-Krikri-8B-Instruct) | 8.29 | 29.87 | 88.13 | 0.895 | 0.695 | - |
|  | [QLoRA](https://huggingface.co/ilsp/llama-krikri-8b-ag-mg-qlora) | 11.90 | 34.07 | 84.16 | 0.906 | **0.713** | +3.60 |
|  | [Full FT](https://huggingface.co/ilsp/llama-krikri-8b-ag-mg-full-ft) | **13.16** | **34.71** | 83.68 | 0.848 | 0.702 | +4.45 |


### Stress Set Results (Rare Dialects)

Evaluated on the 250 sentence-pairs Stress Set (Ionic, Doric, Homeric dialects).

| **Model** | **Method** | **BLEU ↑** | **chrF++ ↑** | **TER ↓** | **BERTScore F1 ↑** | **COMET ↑** | **ΔBLEU** |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **NLLB-600M** | [Base](https://huggingface.co/facebook/nllb-200-distilled-600M) | 0.77 | 14.40 | 118.13 | 0.866 | 0.484 | - |
|  | [LoRA](https://huggingface.co/ilsp/nllb-200-600M-ag-mg-lora) | 5.65 | 28.74 | 88.01 | 0.900 | 0.638 | +4.89 |
| **NLLB-1.3B** | [Base](https://huggingface.co/facebook/nllb-200-distilled-1.3B) | 1.25 | 16.15 | 107.03 | 0.873 | 0.525 | - |
|  | [LoRA](https://huggingface.co/ilsp/nllb-200-1.3B-ag-mg-lora) | 5.68 | 28.94 | 88.24 | 0.900 | 0.656 | +4.43 |
| **M2M100-1.2B** | [Base](https://huggingface.co/facebook/m2m100_1.2B) | 0.07 | 9.37 | 100.34 | 0.840 | 0.427 | - |
| 👉 | [QLoRA](https://huggingface.co/ilsp/m2m100-1.2B-ag-mg-qlora) | 9.52 | 33.30 | 81.95 | **0.911** | 0.691 | **+9.45** |
|  | [Full FT](https://huggingface.co/ilsp/m2m100-1.2B-ag-mg-full-ft) | 8.16 | 31.12 | 83.11 | 0.907 | 0.664 | +8.09 |
| **Krikri-8B-Instruct** | [Base](https://huggingface.co/ilsp/Llama-Krikri-8B-Instruct) | 6.55 | 28.98 | 87.38 | 0.900 | 0.675 | - |
|  | [QLoRA](https://huggingface.co/ilsp/llama-krikri-8b-ag-mg-qlora) | 10.37 | 34.09 | 82.28 | **0.911** | **0.717** | +3.82 |
|  | [Full FT](https://huggingface.co/ilsp/llama-krikri-8b-ag-mg-full-ft) | **12.80** | **35.90** | **81.40** | 0.884 | 0.716 | +6.11 |

---

## Citation

If you use this model, please cite our LREC 2026 paper:

> Mavromatis, S., Sofianopoulos, S., Prokopidis, P., & Giagkou, M. (2026).
> *Ancient Greek to Modern Greek Machine Translation: A Novel Benchmark and
> Fine-Tuning Experiments on LLMs and NMT Models.* In Proceedings of the
> Fifteenth Language Resources and Evaluation Conference (LREC 2026)
> (pp. 8685–8698). European Language Resources Association (ELRA).
> https://doi.org/10.63317/4cdk64dgm2w9

```bibtex
@inproceedings{mavromatis-etal-2026-ancient,
  title     = {Ancient Greek to Modern Greek Machine Translation: A Novel Benchmark and Fine-Tuning Experiments on LLMs and NMT Models},
  author    = {Mavromatis, Spyridon and Sofianopoulos, Sokratis and Prokopidis, Prokopis and Giagkou, Maria},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  month     = {May},
  year      = {2026},
  pages     = {8685--8698},
  address   = {Palma, Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  editor    = {Piperidis, Stelios and Bel, Núria and van den Heuvel, Henk and Ide, Nancy and Krek, Simon and Toral, Antonio},
  doi       = {10.63317/4cdk64dgm2w9}
}
```

**Note on resources:** The fine-tuned models are publicly released. The accompanying AG-MG Parallel Corpus is not publicly distributed due to the complex and uncertain copyright status of the source materials.