Llama-3.1-EstLLM-8B-Instruct-1125

Llama 3.1 EstLLM 8B 1125 Instruct

Llama-3.1-EstLLM-8B-Instruct-1125 is obtained by applying the chat-vector merge approach to tartuNLP/Llama-3.1-EstLLM-8B-Instruct-0825.

The model underwent continuous pre-training starting from meta-llama/Llama-3.1-8B on approximately 35B tokens, which resulted in tartuNLP/Llama-3.1-EstLLM-8B-0525, then supervised fine-tuning and direct preference optimization were applied.

Use with transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "tartuNLP/Llama-3.1-EstLLM-8B-Instruct-1125"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype="auto",
    device_map="auto"
)

# to use on apple silicon, load the following way
# model = AutoModelForCausalLM.from_pretrained(
#     model_name,
#     dtype=torch.float16,
#     device_map="mps",
# )

tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "user", "content": "Kas sa räägid eesti keelt?"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

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

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.4,
    # specify eos token to stop at the end of the assistant response
    eos_token_id=tokenizer.eos_token_id,
)

# generated_ids include the input tokens as well, so we only decode new tokens
response = tokenizer.decode(
    generated_ids[0][model_inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)

print(response)

Model Details

Model Description

• Developed by: TartuNLP and TalTechNLP research groups
• Funded by: Estonian Ministry of Education and Research, “Estonian Language Technology Program 2018-2027”
• Model type: Causal Language Model, Instruction-following
• Language(s) (NLP): Estonian, English
• License: Llama 3.1 Community License Agreement
• Finetuned from model: tartuNLP/Llama-3.1-EstLLM-8B-0525

Continued Pre-Training

Continued Pre-Training was performed for a single epoch on:

• Estonian National Corpus (8.6B tokens)
• Python-Edu (3.3B tokens)
• FineMath4-Plus (9.5B tokens)
• General Instruction-Augmented Corpora (7.4B tokens)
• Cosmopedia v2 (6.9B tokens)

Supervised Fine-Tuning

Approximately 764k examples were used for Supervised Fine-Tuning. The examples mainly come from the Tulu 3 SFT mixture and EuroBlocks. Additional data provided by the Institute of Estonian Language (EKI) was also used. In total about 80% of examples are in English. More details TBA.

Direct Preference Optimization

English-only HelpSteer3 was used as is in the Direct Preference Optimization step, as previous research on Poro 2 models showed that there's no observable benefit from translating preference pairs.

Evaluation

Logits-based

Scores for logits-based evaluation benchmarks are available on the EuroEval leaderboard.

Generative

Every benchmark in this category is treated as a generative problem, and thus the evaluation is performed on the model responses obtained with 0 temperature (not logits). The top scores are higlighted with bold. Second best scores are highlighted with italic bold. Rows are sorted in descending order based on the number of parameters of models (not scores). The arrow up symbol (↑) next to the score indicates an improvement compared to the previous version of the model (Llama-3.1-EstLLM-8B-Instruct-0825). The test set is used for evaluation of each dataset unless noted otherwise.

Note that all models are evaluated with the same prompt template for comparability, meaning that the scores do not necessarily represent each model's best possible performance. This is especially the case for deepseek-ai/DeepSeek-V3-0324 on some of the benchmarks.

Only models of comparable size are evaluated on benchmarks in English.

Instruction-following

Estonian

Instruction level strict accuracy is reported for IFEval-et.

English

Instruction level strict accuracy is reported for IFEval-en.

Multiple Choice

All datasets except Winogrande-et are evaluated in 0-shot mode. Winogrande-et is evaluated in 3-shot mode. Exact match accuracy is reported for every dataset.

Estonian Language Competence

Knowledge and Reasoning (Estonian)

Knowledge and Reasoning (English)

Translation

English to Estonian

Limitations

This is an early prototype version. Accordingly, it has limitations in addition to the base Llama limitations:

• Relatively short context of 4096 tokens. It's not expected to perform well on context sizes beyond that. Merging somewhat mitigates that.
• Multi-turn conversations are not guaranteed to be supported, although this capability is improved by merging.
• Trained with the original Llama 3.1 system prompt that has a hard-coded date cut-off.
• There may be unexpected side effects as a result of merging.

Citation

@misc{dorkin2026estllmenhancingestoniancapabilities,
      title={{EstLLM: Enhancing Estonian Capabilities in Multilingual LLMs via Continued Pretraining and Post-Training}}, 
      author={Aleksei Dorkin and Taido Purason and Emil Kalbaliyev and Hele-Andra Kuulmets and Marii Ojastu and Mark Fišel and Tanel Alumäe and Eleri Aedmaa and Krister Kruusmaa and Kairit Sirts},
      year={2026},
      eprint={2603.02041},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.02041}, 
}