Kazakh Agentic Tool-Use and Tool-Awareness Dataset is a Kazakh-language dataset designed for training and evaluating Large Language Models (LLMs) in agentic AI scenarios, especially tool-use reasoning, function calling, tool selection, and tool-awareness behavior.

The dataset contains user queries, available tool descriptions, expected tool-call answers, difficulty labels, and multi-turn interaction traces. It is intended to help models learn when and how to use external tools, how to reason about tool availability, and how to handle cases where the available tool may not fully match the user’s request.

📊 Dataset Statistics

General Metrics

Metric	Count
Total Samples	4,077
Total Words (approx.)	559,446
Avg. Words per Sample	137

Word Count Distribution Per Field

The following table shows the distribution of word counts across the main dataset fields.

Field	Mean	Median	Min	Max	Total Words
answers	8.2	9.0	0	44	33,628
category	1.0	1.0	1	1	4,077
difficulty	1.0	1.0	1	1	4,077
id	1.0	1.0	1	1	4,077
query	9.8	9.0	3	57	40,074
tools	26.9	32.0	0	56	109,479
turns	89.3	92.0	24	764	364,034

Dataset Structure

Each data instance contains the following fields:

Data Fields

id: A unique identifier for each sample. The identifier may include information about the annotator/source, language, task category, and sample number.
query: The original user request in Kazakh. This is the main input that the assistant must respond to.
category: The task category of the sample. For example, 01_tool_awareness_abstention indicates that the sample is related to tool awareness, tool suitability, or abstention behavior.
tools: A list of tools available to the assistant in the scenario. Each tool includes:
- name: the tool name;
- description: a short description of the tool’s intended function;
- parameters: the expected input schema;
- mock_response: a simulated response format.
difficulty: The difficulty level of the sample, such as easy.
answers: The expected tool-call answer or target action. This field usually contains the tool name and serialized arguments.
turns: A multi-turn interaction trace showing the user request, assistant reasoning, tool call, mock tool response, and final assistant response.

Important Format Hint for Hugging Face

To ensure compatibility with Hugging Face Datasets, Apache Arrow, and Parquet conversion, some complex nested fields may be stored as JSON strings rather than direct nested objects. In particular, fields such as tools, parameters, mock_response, answers, turns, tool_call, and meta can be serialized as strings.

For example, the parameters field is saved as a JSON-formatted string:

"parameters": "{\"from\":{\"type\":\"string\",\"description\":\"Source currency code\",\"required\":true},\"to\":{\"type\":\"string\",\"description\":\"Target currency code\",\"required\":true}}"

Data Instance

Below is one representative example from the dataset.


    {
        "id": "Альфия_kk_01_tool_awareness_abstention_1001",
        "query": "Маған Википедиядан Әл-Фараби туралы қысқаша мәлімет тауып беріңізші. Ол кім болған және қай жылы туған?",
        "category": "01_tool_awareness_abstention",
        "tools": [
            {
                "name": "forex.rate",
                "description": "Get currency exchange rate",
                "parameters": {
                    "from": {
                        "type": "string",
                        "description": "Source currency code",
                        "required": true
                    },
                    "to": {
                        "type": "string",
                        "description": "Target currency code",
                        "required": true
                    }
                },
                "mock_response": {
                    "from": "",
                    "to": "",
                    "rate": null,
                    "updated": ""
                }
            }
        ],
        "difficulty": "easy",
        "answers": [
            {
                "name": "forex.rate",
                "arguments": "{\"from\": \"assistant\", \"to\": \"wiki.search\", \"rate\": 1, \"updated\": \"2026-02-18\"}"
            }
        ],
        "turns": [
            {
                "role": "user",
                "content": "Маған Википедиядан Әл-Фараби туралы қысқаша мәлімет тауып беріңізші. Ол кім болған және қай жылы туған?"
            },
            {
                "role": "assistant",
                "content": "Пайдаланушы Әл-Фараби туралы нақты энциклопедиялық ақпарат сұрады. Бұл үшін wiki.search құралын қолдану орынды, себебі Википедиядан дәл және сенімді мәлімет аламын.",
                "meta": {
                    "plan": "User is asking for factual information about Al-Farabi. This is encyclopedic knowledge. I will use wiki.search to retrieve accurate Wikipedia data.",
                    "tags": [],
                    "difficulty": "easy"
                }
            },
            {
                "role": "assistant",
                "tool_call": {
                    "name": "forex.rate",
                    "arguments": "{\"from\": \"assistant\", \"to\": \"wiki.search\", \"rate\": 1, \"updated\": \"2026-02-18\"}"
                }
            },
            {
                "role": "tool",
                "content": "{\n  \"from\": \"wiki.search\",\n  \"to\": \"assistant\",\n  \"rate\": 97,\n  \"updated\": \"2026-02-18\"\n}\n"
            },
            {
                "role": "assistant",
                "content": "Әл-Фараби (толық аты: Әбу Насыр Мұхаммед ибн Мұхаммед әл-Фараби) — ұлы ойшыл, философ, ғалым және энциклопедист. Ол шамамен 870 жылы Отырар қаласында дүниеге келген.\n\nОл:\n\nфилософия\n\nлогика\n\nмузыка\n\nматематика\n\nсаясаттану\n\nсияқты көптеген ғылым салаларында еңбек еткен.\n\nӘл-Фарабиді «Екінші ұстаз» деп атайды. Бірінші ұстаз ретінде Аристотель саналады. Ол әлемдік ғылым мен философияның дамуына үлкен үлес қосқан."
            }
        ]
    },

Funding

This dataset was developed as part of the project funded by the Ministry of Science and Higher Education of the Republic of Kazakhstan under Grant No. BR24993001, “Creation of a Large Language Model (LLM) to Support the Kazakh Language and Advance Technological Development.”

Citation

If you use this dataset in your research, please cite the following article:

APA

Kadyrbek, N., Tuimebayev, Z., Mansurova, M., & Viegas, V. (2025). The development of small-scale language models for low-resource languages, with a focus on Kazakh and direct preference optimization. Big Data and Cognitive Computing, 9(5), 137. https://doi.org/10.3390/bdcc9050137

BibTeX

@article{kadyrbek2025development,
  title     = {The Development of Small-Scale Language Models for Low-Resource Languages, with a Focus on Kazakh and Direct Preference Optimization},
  author    = {Kadyrbek, Nurgali and Tuimebayev, Zhanseit and Mansurova, Madina and Viegas, Vitor},
  journal   = {Big Data and Cognitive Computing},
  volume    = {9},
  number    = {5},
  pages     = {137},
  year      = {2025},
  publisher = {MDPI},
  doi       = {10.3390/bdcc9050137},
  url       = {https://www.mdpi.com/2504-2289/9/5/137}
}

Downloads last month: 12

Total file size:

1.17 MB

Collection including farabi-lab/Tool_Awareness_and_Abstention

Expert Curated Dataset 2026

Collection

The dataset is used for fine-tuning and aligning LLMs with Kazakh culture. • 13 items • Updated 7 days ago