license: llama3.1
language:
- el
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- text-generation-inference
base_model:
- ilsp/Llama-Krikri-8B-Base
🚨 PLEASE USE THE OFFICIAL QUANTIZED VERSIONS: GGUF OR REQUEST A SPECIFIC ONE 🚨
🚨 There is no guarantee that you are using the latest improved versions from 3rd party quantizations as we have updated the model's weights 🚨
Llama-Krikri-8B-Instruct: An Instruction-tuned Large Language Model for the Greek language
Following the release of Meltemi-7B on the 26th March 2024, we are happy to welcome Krikri to the family of ILSP open Greek LLMs. Krikri is built on top of Llama-3.1-8B, extending its capabilities for Greek through continual pretraining on a large corpus of high-quality and locally relevant Greek texts. We present Llama-Krikri-8B-Instruct, along with the base model, Llama-Krikri-8B-Base
Model Information
Base Model
- Vocabulary extension of the Llama-3.1 tokenizer with Greek tokens
- 128k context length (approximately 80,000 Greek words)
- We extend the pretraining of Llama-3.1-8B with added proficiency for the Greek language, by utilizing a large training corpus.
- This corpus includes 56.7 billion monolingual Greek tokens, constructed from publicly available resources.
- Additionaly, to mitigate catastrophic forgetting and ensure that the model has bilingual capabilities, we use additional sub-corpora with monolingual English texts (21 billion tokens) and Greek-English parallel data (5.5 billion tokens).
- The training corpus also contains 7.8 billion math and code tokens.
- This corpus has been processed, filtered, and deduplicated to ensure data quality and is outlined below:
| Sub-corpus | # Tokens | Percentage |
|---|---|---|
| Greek | 56.7 B | 62.3 % |
| English | 21.0 B | 23.1 % |
| Parallel | 5.5 B | 6.0 % |
| Math/Code | 7.8 B | 8.6 % |
| Total | 91 B | 100% |
Chosen subsets of the 91 billion corpus were upsampled resulting in a size of 110 billion tokens.
Instruct Model
Llama-Krikri-8B-Instruct is the result of post-training Llama-Kriki-8B-Base and features:
- Enhanced chat capabilities and instruction-following in both Greek and English.
- Document translation from Greek to English, French, German, Italian, Portuguese, Spanish and vice versa.
- Great performance on generation, comprehension, and editing tasks, such as summarization, creative content creation, text modification, entity recognition, sentiment analysis, etc.
- Domain-specifc expertise for specialized legal, financial, medical, and scientific applications.
- Retrieval-Augmented Generation (RAG) utilizing multiple documents with 128k context length.
- Improved coding and agentic capabilities with correct formatting and tool use.
- Conversion or structured extraction (e.g., XML, JSON) in data-to-text & text-to-data settings.
- Analytical thinking and Chain-of-Thought (CoT) reasoning for problem-solving.
Post-training Methodology
We used a multi-stage process in order to build Llama-Krikri-8B-Instruct which includes:
- 2-stage Supervised Fine-Tuning with a combination of Greek & English instruction-response pairs (& multi-turn conversations)
- Stage 1: 856,946 instruction-response pairs (371,379 Greek + 485,567 English)
- Stage 2: 638,408 instruction-response pairs (279,948 Greek + 358,460 English)
- Alignment with a combination of Greek & English preference triplets (Instruction - Chosen Response - Rejected Response)
- Length Normalized DPO: 92,394 preference triplets (47,132 Greek + 45,262 English)
Post-training Data Construction
To build the SFT & DPO data, we utilized various methodologies including:
- Collecting existing high-quality datasets such as Tulu 3, SmolTalk, MAGPIE Ultra, Orca Agent Instruct, IFEval Like Data, UltraFeedback, NVIDIA HelpSteer2, Intel Orca, UltraMedical, and other datasets focused on safety, truthfulness, and instruction-following.
- Translating various data into Greek using an in-house translation tool.
- Regenerating translated data and contrasting the translated with the regenerated responses (i.e., for creating preference triplets).
- Distilling (with the MAGPIE methodology) models which exhibit strong performance in Greek, such as Gemma 2 27B IT.
- Scoring data with the Skywork Reward Gemma 2 27B v0.2 Reward Model and filtering using rule-based filters.
- Creating data for sentence and document translation using high-quality parallel corpora mainly from ELRC-SHARE.
- Synthetically extracting question-answer pairs and multi-turn dialogues from diverse sources such as Wikipedia, EUR-LEX, Greek School Books, and Kallipos.
Evaluation
In the table below, we report the scores for Greek IFEval (strict) and English IFEval (strict) for various chat models that exhibit strong performance.
We can observe that Llama-Krikri-8B-Instruct exhibits the strongest performance in instruction following for both Greek and English across all the models we tested. In particular, it surpasses Llama-3.1-8B-Instruct by +21.7% and +7.3% on the Greek and English IFEval respectively.
| IFEval EL (strict) | IFEval EN (strict) | |
|---|---|---|
| Qwen 2.5 7B Instruct | 46.2% | 74.8% |
| EuroLLM 9B Instruct | 51.3% | 64.5% |
| Aya Expanse 8B | 50.4% | 62.2% |
| Meltemi 7B v1.5 Instruct | 32.7% | 41.2% |
| Llama-3.1-8B Instruct | 45.8% | 75.1% |
| Llama-Krikri-8B Instruct | 67.5% | 82.4% |
We also used the Arena-Hard-Auto automatic evaluation tool, as well the translated (and post-edited) version for Greek that is publicly available here.
Below, we show the scores for the Greek version of Arena-Hard-Auto for various open and closed chat models that were determined using gpt-4o-2024-08-06 as the judge model and gpt-4o-mini-2024-07-18 as the baseline model (i.e., by default 50% score).
Please note that recent research has shown that judge models are biased towards student models, i.e., models finetuned on distilled data from a stronger/larger teacher model. While post-training data of GPT-4o-Mini are undisclosed, it would be very reasonable to assume that it has been trained -at least partly- with GPT-4o serving as the teacher model and therefore that the judge is biased towards the baseline model.
Below, we show the scores for the original Arena-Hard-Auto dataset for various open and closed chat models. We followed the original methodology of using gpt-4-1106-preview as the judge model and gpt-4-0314 as the baseline model.
🚨 More information on post-training, methodology, and evaluation coming soon. 🚨
How to use
With Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda"
model = AutoModelForCausalLM.from_pretrained("ilsp/Llama-Krikri-8B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("ilsp/Llama-Krikri-8B-Instruct")
model.to(device)
system_prompt = "Είσαι το Κρικρί, ένα εξαιρετικά ανεπτυγμένο μοντέλο Τεχνητής Νοημοσύνης για τα ελληνικα και εκπαιδεύτηκες από το ΙΕΛ του Ε.Κ. \"Αθηνά\"."
user_prompt = "Σε τι διαφέρει ένα κρικρί από ένα λάμα;"
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
input_prompt = tokenizer(prompt, return_tensors='pt').to(device)
outputs = model.generate(input_prompt['input_ids'], max_new_tokens=256, do_sample=True)
print(tokenizer.batch_decode(outputs)[0])
With OpenAI compatible server via vLLM
vllm serve ilsp/Llama-Krikri-8B-Instruct \
--enforce-eager \
--dtype 'bfloat16' \
--api-key token-abc123
Then, the model can be used through Python using:
from openai import OpenAI
api_key = "token-abc123"
base_url = "http://localhost:8000/v1"
client = OpenAI(
api_key=api_key,
base_url=base_url,
)
system_prompt = "Είσαι ένα ανεπτυγμένο μεταφραστικό σύστημα που απαντάει με λίστες Python. Δεν γράφεις τίποτα άλλο στις απαντήσεις σου πέρα από τις μεταφρασμένες λίστες."
user_prompt = "Δώσε μου την παρακάτω λίστα με μεταφρασμένο κάθε string της στα ελληνικά: ['Ethics of duty', 'Postmodern ethics', 'Consequentialist ethics', 'Utilitarian ethics', 'Deontological ethics', 'Virtue ethics', 'Relativist ethics']"
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
]
response = client.chat.completions.create(model="ilsp/Llama-Krikri-8B-Instruct",
messages=messages,
temperature=0.0,
top_p=0.95,
max_tokens=8192,
stream=False)
print(response.choices[0].message.content)
# ['Ηθική καθήκοντος', 'Μεταμοντέρνα ηθική', 'Συνεπειοκρατική ηθική', 'Ωφελιμιστική ηθική', 'Δεοντολογική ηθική', 'Ηθική αρετών', 'Σχετικιστική ηθική']
Evaluation
🚨 Instruction following and chat capability evaluation benchmarks coming soon. 🚨
Acknowledgements
The ILSP team utilized Amazon's cloud computing services, which were made available via GRNET under the OCRE Cloud framework, providing Amazon Web Services for the Greek Academic and Research Community.