AJAssel Jaxylykova

Assel Jaxylykova

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Assel Jaxylykova has a total of 15 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Noise-Aware Direct Preference Optimization for RLAIF

    by Alymzhan Toleu, Gulmira Tolegen, Alexandr Pak, Assel Jaxylykova

    Reinforcement Learning from Human Feedback (RLHF) produces powerful instruction-following models but relies on a preference-labeling process that is both costly and slow. An effective alternative, Reinforcement Learning from AI Feedback (RLAIF), uses large language models as teachers for relabeling; however, this introduces substantial label noise. In our setting, we found that AI teachers flipped approximately 50% of the original human preferences on the dataset, a condition that degrades the performance of standard direct preference optimization (DPO). We propose noise-robust DPO (nrDPO) and nrDPO-gated, two drop-in variants that make DPO resilient to noisy preferences. nrDPO reweights each pair by (i) a margin-confidence term from a frozen reference policy (base or SFT), (ii) a context-stability term that penalizes preferences that change under truncated histories, and (iii) a length correction to curb verbosity bias. nrDPO-gated further filters low-confidence pairs via a simple threshold on the reference margin. On a dataset with heavy synthetic noise (30% flips), nrDPO-gated improves the preference accuracy by +3.8% over vanilla DPO; in a realistic RLAIF setting, nrDPO-gated is the only configuration that recovers competitive alignment, reaching ≈60% on a 5k relabeled set (vs. ≈49–50% for vanilla DPO) and approaching RLHF baselines.


  • Open Access

    ADAPTATION OF TEXT GENERATION STYLE TO A SPECIFIC AUDIENCE OR CONTENT

    by Zh. Zhangbyrbay, I. Akhmetov, A. Pak, A. Jaxylykova, P. Komada, А. Джаксылыкова

    Adaptation of text generation style to specific audiences or content can be achieved without costly fine-tuning. We freeze model weights and instead (i) search eight decoder hyperparameters with Bayesian optimization and (ii) prepend a one-line style cue that modulates readability. Experiments on five mathematical question-answering benchmarks (AQUA-RAT, MathQA, GSM8K, MAWPS, SVAMP) with three 8–14 B-parameter checkpoints (LLaMA-3.1-8B, DeepSeek-Qwen-8B/14B) show that 50-trial Optuna searches raise exact-match accuracy by up to 36 percentage points and close 5–10 points of the gap to 30–70 B fine-tuned baselines. The same settings transfer across tasks with under 2-point loss. Adding the child-friendly header leaves accuracy virtually unchanged while halving the Flesch–Kincaid grade level and shortening reasoning traces. All experiments fit within a few GPU-hours on a single A100, making the method practical for resource-constrained deployments. The study demonstrates that careful decoder control combined with micro-prompts delivers numerical correctness and audience-appropriate exposition without additional training or tuning time.

  • Open Access

    End-to-End Multi-Modal Speaker Change Detection with Pre-Trained Models

    by Alymzhan Toleu, Gulmira Tolegen, Alexandr Pak, Jaxylykova Assel, Bagashar Zhumazhanov

    In this work, we propose a multi-modal speaker change detection (SCD) approach with focal loss, which integrates both audio and text features to enhance detection performance. The proposed approach utilizes pre-trained large-scale models for feature extraction and incorporates a self-attention mechanism to optimize useful features related to speaker change. The extracted features are fused and processed through a fully connected classification network, with layer normalization and dropout for stability and generalization. To address class imbalance, we apply focal loss, which reduces errors for the difficult samples, leading to better balanced performance. Extensive experiments on a multi-talker meeting dataset demonstrate that the proposed multi-modal approach consistently outperforms single-modal models, proving the complementary nature of audio and text for SCD. Fine-tuning pre-trained models (Wav2Vec2 and Bert) for audio and text significantly boosts accuracy, achieving a 21% improvement over frozen models. The self-attention mechanism further improves performance by 2%, highlighting its ability to capture speaker transition cues effectively. Additionally, focal loss enhances the model’s performance, making it more robust to imbalanced data.


  • Open Access

    End-to-End Multi-Modal Speaker Change Detection with Pre-trained Models

    by Alymzhan Toleu, Gulmira Tolegen, Alexandr Pak, Assel Jaxylykova, Bagashar Zhumazhanov

    In this work, we propose a multi-modal Speaker Change Detection (SCD) approach with focal loss that integrates both audio and text features to enhance detection performance. % The proposed approach utilizes pre-trained large-scale models for feature extraction and incorporates a self-attention mechanism to optimize the useful features related to speaker change. % The extracted features are fused and processed through a fully connected classification network, with layer normalization and dropout for stability and generalization. % To address class imbalance, we apply focal loss, which reduces errors for the difficult samples, leading to better balanced performance. Extensive experiments on a multi-talker meeting dataset demonstrate that the proposed multi-modal approach consistently outperforms single-modal models, proving the complementary nature of audio and text for SCD. % Fine-tuning pre-trained models (Wav2Vec2 and Bert) for audio and text significantly boosts accuracy, achieving a 21\% improvement over frozen models. % The self-attention mechanism further improves performance by 2\%, highlighting its ability to capture speaker transition cues effectively. % Additionally, focal loss enhances model's performance, making it more robust to imbalanced data.


  • Open Access

    Using natural language processing techniques to analyze the use of Google Earth Engine in 2015–2023 remote sensing research publications

    by Institute of Information and Computational Technologies, A.B. Jaxylykova, A.M. Mirash, BL-group, A.A. Pak, Armanovich Ziyaden

  • SCIENTIFIC NAMED ENTITY RECOGNITION WITH THE HELP OF MODERN METHODS

    by A.M. Yelenov, A.B. Jaxylykova

    This research focuses on a comparative study of the Named Entity Recognition task for scientific article texts. Natural language processing could be considered as one of the cornerstones in the machine learning area which devotes its attention to the problems connected with the understanding of different natural languages and linguistic analysis. It was already shown that current deep learning techniques have a good performance and accuracy in such areas as image recognition, pattern recognition, computer vision, that could mean that such technology probably would be successful in the neuro-linguistic programming area too and lead to a dramatic increase on the research interest on this topic. For a very long time, quite trivial algorithms have been used in this area, such as support vector machines or various types of regression, basic encoding on text data was also used, which did not provide high results. The following dataset was used to process the experiment models: Dataset Scientific Entity Relation Core. The algorithms used were Long short-term memory, Random Forest Classifier with Conditional Random Fields, and Named-entity recognition with Bidirectional Encoder Representations from Transformers. In the findings, the metrics scores of all models were compared to each other to make a comparison. This research is devoted to the processing of scientific articles, concerning the machine learning area, because the subject is not investigated on enough properly level.The consideration of this task can help machines to understand natural languages better, so that they can solve other neuro-linguistic programming tasks better, enhancing scores in common sense.