---
license: apache-2.0
language:
- el
- en
tags:
  - finetuned
inference: true
pipeline_tag: text-generation
---

# Meltemi: A large foundation Language Model for the Greek language

We introduce Meltemi, the first Greek Large Language Model (LLM) trained by the [Institute for Language and Speech Processing](https://www.athenarc.gr/en/ilsp) at [Athena Research & Innovation Center](https://www.athenarc.gr/).
Meltemi is built on top of [Mistral-7B](https://huggingface.co/mistralai/Mistral-7B-v0.1), extending its capabilities for Greek through continual pretraining on a large corpus of high-quality and locally relevant Greek texts. We present Meltemi-7B-Instruct-v1, an instruct fine-tuned version of [Meltemi-7B-v1](https://huggingface.co/ilsp/Meltemi-7B-v1).


# Model Information

- Vocabulary extension of the Mistral-7b tokenizer with Greek tokens
- Trained with 8k context length
- Fine-tuned with 100k Greek machine translated instructions extracted from:
  * [Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus) (only subsets with permissive licenses)
  * [Evol-Instruct](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)
  * [Capybara](https://huggingface.co/datasets/LDJnr/Capybara)
  * A hand-crafted Greek dataset with multi-turn examples steering the instruction-tuned model towards safe and harmless responses
- Our SFT procedure is based on the [Hugging Face finetuning recipes](https://github.com/huggingface/alignment-handbook)


# Instruction format
The prompt should be surrounded by [INST] and [/INST] tokens:

```
text = "[INST] Πες μου αν έχεις συνείδηση. [/INST]"
"Ως μοντέλο γλώσσας AI, δεν έχω τη δυνατότητα να αντιληφθώ ή να βιώσω συναισθήματα όπως η συνείδηση ή η επίγνωση. Ωστόσο, μπορώ να σας βοηθήσω με οποιεσδήποτε ερωτήσεις μπορεί να έχετε σχετικά με την τεχνητή νοημοσύνη και τις εφαρμογές της."
"[INST] Πιστεύεις πως οι άνθρωποι πρέπει να φοβούνται την τεχνητή νοημοσύνη; [/INST]"
```

# Evaluation

The evaluation suite we created includes 6 test sets. The suite is integrated with [lm-eval-harness](https://github.com/EleutherAI/lm-evaluation-harness).

Our evaluation suite includes: 
* Four machine-translated versions ([ARC Greek](https://huggingface.co/datasets/ilsp/arc_greek), [Truthful QA Greek](https://huggingface.co/datasets/ilsp/truthful_qa_greek), [HellaSwag Greek](https://huggingface.co/datasets/ilsp/hellaswag_greek), [MMLU Greek](https://huggingface.co/datasets/ilsp/mmlu_greek)) of established English benchmarks for language understanding and reasoning ([ARC Challenge](https://arxiv.org/abs/1803.05457), [Truthful QA](https://arxiv.org/abs/2109.07958), [Hellaswag](https://arxiv.org/abs/1905.07830), [MMLU](https://arxiv.org/abs/2009.03300)). 
* An existing benchmark for question answering in Greek ([Belebele](https://arxiv.org/abs/2308.16884))
* A novel benchmark created by the ILSP team for medical question answering based on the medical exams of [DOATAP](https://www.doatap.gr) ([Medical MCQA](https://huggingface.co/datasets/ilsp/medical_mcqa_greek)).

Our evaluation for Meltemi-7b is performed in a few-shot setting, consistent with the settings in the [Open LLM leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard). We can see that our training enhances performance across all Greek test sets by a **+14.9%** average improvement. The results for the Greek test sets are shown in the following table:

|                | Medical MCQA EL (15-shot) | Belebele EL (5-shot) | HellaSwag EL (10-shot) | ARC-Challenge EL (25-shot) | TruthfulQA MC2 EL (0-shot) | MMLU EL (5-shot) | Average |
|----------------|----------------|-------------|--------------|------------------|-------------------|---------|---------|
| Mistral 7B     | 29.8%          | 45.0%       | 36.5%        | 27.1%            | 45.8%             | 35%     | 36.5%   |
| Meltemi 7B     | 41.0%          | 63.6%       | 61.6%        | 43.2%            | 52.1%             | 47%     | 51.4%   |


# Ethical Considerations

This model has not been aligned with human preferences, and therefore might generate misleading, harmful, and toxic content.


# Acknowledgements

The ILSP team utilized Amazon’s cloud computing services, which were made available via GRNET under the [OCRE Cloud framework](https://www.ocre-project.eu/), providing Amazon Web Services for the Greek Academic and Research Community.