News

KazLLM, developed by Kazakhstan's ISSAI institute at Nazarbayev University, received recognition from a highly respected figure: “At the AI Action Summit in Paris, I met a delegation from Kazakhstan that showed me KazLLM—a version of Llama adapted for Kazakh and other languages spoken in the region. Excellent results in Kazakh.”

Sincere congratulations from the DSML KZ community to everyone involved!

LinkedIn post ISSAI model

• Yann LeCun is a French AI scientist and one of the founders of the field of deep learning
• He pioneered neural networks and machine-learning algorithms, including convolutional neural network architecture (CNN)
• He currently works at Facebook AI Research (FAIR) and is a professor at New York University

• MMLU in Kazakh: Evaluates multitask language understanding
• GSM8K in Kazakh: Tests mathematical reasoning
• Kazakh Constitution MC: Tests legal and civic knowledge
• Kazakh Dastur MC: Evaluates understanding of cultural heritage and customs
• Unified National Test MC: Covers history, literature, biology, and other subjects

We are happy to share that the first model-comparison results are now available!

• OpenAI's general-purpose GPT-4o confidently holds first place among all models
• In the open-source segment, Kazakhstan's ISSAI laboratory leads with LLaMA-3.1-KazLLM-1.0-8B

Next steps:

• Release an offline arena with ELO ratings
• Add more challenging benchmarks for real-world tasks
• Release custom instruction datasets

More details:

• View the leaderboard
• Submit your model

The team welcomes feedback and invites everyone interested to collaborate!

Community members Sanzhar Murzakhmetov @stringersolo, Beksultan Sagyndyk @bsagyndyk, and Kirill Yakunin @JenarosHat released a series of benchmarks for language understanding, reasoning, and other features of Kazakh.

1) The team translated several popular datasets into Kazakh:

• MMLU in Kazakh: Evaluates multitask language understanding
• GSM8K in Kazakh: Tests mathematical reasoning

2) They created custom Kazakhstan-focused multiple-choice benchmarks:

• Kazakh Constitution MC: Tests legal and civic knowledge
• Kazakh Dastur MC: Evaluates understanding of cultural heritage and customs
• Unified National Test MC: Covers history, literature, biology, and other subjects

The team is building an open leaderboard for these benchmarks and a platform for testing and comparing models developed for Kazakh.

Read more in the team's LinkedIn post.

Zangir studied how neural networks and reinforcement learning can accelerate solutions to combinatorial problems such as routing, resource allocation, and planning. He presented the following projects:

• A new training approach for production planning in which problem complexity increases dynamically, improving model training
• An RL framework for routing problems with stochastic travel times on a graph
• A new self-directed search strategy that uses large language models to solve combinatorial problems efficiently

This work was published at IJCAI23, ACML23, IEEE Access, and NeurIPS24.

MBZUAI is the world's first university dedicated to artificial intelligence. It opened in Abu Dhabi in 2019 and ranks among the top 20 in AI/ML. Zangir became the third PhD candidate in MBZUAI's history to successfully defend his dissertation!

Applications recently opened for MBZUAI Master's and PhD programs starting in fall 2025. The university offers five areas: ML, CV, NLP, CS, and Robotics. The first application-round deadline is January 15, 2025.