<p align="center">
  <a href="https://huggingface.co/datasets/cisimcik/turkish-chat-max-25k/blob/main/README_en.md"><img src="https://img.shields.io/badge/English-2563eb?style=for-the-badge" alt="English"></a>
  <a href="https://huggingface.co/datasets/cisimcik/turkish-chat-max-25k"><img src="https://img.shields.io/badge/T%C3%BCrk%C3%A7e-6b7280?style=for-the-badge" alt="Türkçe"></a>
</p>

# Turkish Chat Max 25K

**24,543 synthetic Turkish instruction-tuning conversations**: 16 task types, step-by-step reasoning when a turn
needs it, multi-turn dialogue and tool calling. Every row was produced by our own data generator, and the dataset
contains **no text copied from existing datasets**.

| | |
|---|---|
| Conversations | 24,543 |
| Multi-turn (2–5 user turns) | 5,160 (21.0%) |
| With reasoning (per-turn rationale) | 18,056 (73.6%) |
| With tool calls | 921 rows, 1,307 calls |
| Languages | Turkish; English appears in translation and English-source tasks |
| Format | Parquet, Hugging Face chat layout (`messages`) |
| License | Apache 2.0 (free for any use, including commercial use) |

## Quick start
```python
from datasets import load_dataset

ds = load_dataset("cisimcik/turkish-chat-max-25k", split="train")                       # everything (config "all")
tools = load_dataset("cisimcik/turkish-chat-max-25k", "tool_calling", split="train")    # one task type
print(ds[0]["messages"])
```
Every task type is also its own config, so it can be browsed separately in the viewer. The configs are
`extractive_qa`, `grammatical_error_correction`, `multiple_choice_qa`, `summarization`, `open_qa`, `natural_language_inference`, `machine_translation`, `math_and_logic`, `text_classification`, `instruction_following`, `code`, `tool_calling`, `information_extraction`, `casual_conversation`, `creative_writing`, `planning`.

Filter on the metadata columns:
```python
hard_math = ds.filter(lambda r: r["task_type"] == "Mathematical & Logical Reasoning" and r["difficulty"] in ("hard", "expert"))
no_thinking = ds.filter(lambda r: not r["has_reasoning"])
dialogues = ds.filter(lambda r: r["n_turns"] >= 3)
```
With pandas or polars, read the Parquet files directly: `pl.read_parquet("hf://datasets/cisimcik/turkish-chat-max-25k/data/*/train.parquet")`.

## Data format
One row is one conversation. `messages` is a list of turns:

| field | meaning |
|---|---|
| `role` | `system`, `user`, `assistant` or `tool` |
| `content` | the visible text of the turn |
| `reasoning` | the assistant's rationale for this turn (null when the turn did not think) |
| `tool_calls` | assistant turns that call tools: list of `{name, arguments}` (`arguments` is a JSON string) |
| `name` | on `tool` turns: the tool that produced this result |

Columns for filtering and sorting:

| column | values |
|---|---|
| `task_type`, `task_subtype` | the task (table below) and its variant, e.g. `Extractive QA` / `howto_span`, `Tool Use & Function Calling` / `parallel_calls` |
| `difficulty` | easy, medium, hard, expert |
| `reasoning_effort` | deepest reasoning level in the row: none, low, medium, high, xhigh |
| `reasoning_mode` | `effort` (reasoning level given), `auto` (the assistant decides per turn, including not thinking), `off` (no reasoning) |
| `has_reasoning`, `n_turns`, `has_system`, `has_tools`, `n_tools` | quick filters |
| `tools` | JSON string of the function schemas offered in the conversation (tool rows) |
| `language_mode`, `prompt_language` | `tr_only`, `tr_en` (Turkish with English material), `en_source`; and the language of the user prompt |
| `culture`, `topic` | cultural frame and topic the conversation was built on |
| `user_prompt_style`, `register` | how the user writes (informal, typos, terse, exam, polite, …) and the formality |
| `answer_style`, `source_format`, `interaction`, `challenge` | expected answer shape, source layout (paragraph, dialogue, bullets, table), multi-turn pattern, difficulty device (distractors, negation, boundary case) |
| `answerability`, `abstain_kind` | `answerable` / `unanswerable` and why (false premise, missing information, future event, …) |
| `id`, `recipe_version`, `created` | row id, generator recipe version, creation date |

A tool-calling row (texts shortened):
```json
{
  "id": "0040b31e-704e-4f02-b6dd-fed8372e70df",
  "task_type": "Tool Use & Function Calling",
  "task_subtype": "sequential_calls",
  "difficulty": "hard",
  "reasoning_effort": "low",
  "n_turns": 2,
  "has_tools": true,
  "messages": [
    {
      "role": "system",
      "content": "Sen yardımcı bir asistanısın. Gerektiğinde tanımlı araçları çağır. Zorunlu bir bilgi eksikse tahmin etme, sor; araç sonucunda olmayan bir şeyi uydurma. Aksi ist…"
    },
    {
      "role": "user",
      "content": "sipariş no YL-4821. marketing için e-posta ve SMS iznimi kapat; sipariş haberlerini kesme. Teslimat adresimle diğer sipariş bilgilerimi de silme, paket yolda. B…"
    },
    {
      "role": "assistant",
      "content": "",
      "tool_calls": [
        {
          "name": "siparis_kaydi_getir",
          "arguments": "{\"siparis_no\": \"YL-4821\"}"
        }
      ]
    },
    {
      "role": "tool",
      "content": "{\"siparis_no\":\"YL-4821\",\"musteri_id\":\"M-731\",\"saklanan_kisisel_veri_alanlari\":[\"ad soyad\",\"e-posta adresi\",\"telefon numarası\",\"teslimat adresi\"],\"pazarlama_izin…",
      "name": "siparis_kaydi_getir"
    },
    {
      "role": "assistant",
      "content": "",
      "reasoning": "Kayıttaki müşteri kimliği M-731; iki pazarlama iznini bu kimlikle kapatmalıyım. Güncelleme sipariş bilgilerine ve işlem bildirimlerine dokunmadığı için korunmas…",
      "tool_calls": [
        {
          "name": "pazarlama_tercihlerini_guncelle",
          "arguments": "{\"musteri_id\": \"M-731\", \"eposta_izni\": false, \"sms_izni\": false}"
        }
      ]
    },
    {
      "role": "tool",
      "content": "{\"musteri_id\":\"M-731\",\"basarili\":true,\"pazarlama_izinleri\":{\"eposta\":false,\"sms\":false},\"siparis_bildirimleri\":true,\"siparis_bilgileri_korundu\":true,\"saklanan_k…",
      "name": "pazarlama_tercihlerini_guncelle"
    },
    {
      "role": "assistant",
      "content": "E-posta ve SMS pazarlama izinlerin kapatıldı. Sipariş bildirimlerin açık kaldı; teslimat adresin ve diğer sipariş bilgilerin silinmedi.\n\nTutulan kişisel bilgi a…"
    },
    {
      "role": "user",
      "content": "başka bi hesap var: Derya, Cansu ve ben yemek masrafını eşit paylaşacağız. toplam 840 lira, Derya 500 Cansu 340 ödedi ben hiç ödemedim. şimdi kime kaç lira gönd…"
    },
    {
      "role": "assistant",
      "content": "Derya’ya **220 TL**, Cansu’ya **60 TL** gönder. Böylece üçünüz de kişi başı **280 TL** ödemiş olursunuz.",
      "reasoning": "Toplam 840 lirayı üç kişiye bölünce kişi başına 280 lira buluyorum. Derya 500−280=220 lira, Cansu 340−280=60 lira fazla ödüyor. Hiç ödeme yapmayan kişinin 280 l…"
    }
  ],
  "tools": "[\"{\\\"type\\\": \\\"function\\\", \\\"function\\\": {\\\"name\\\": \\\"siparis_kaydi_getir\\\", \\\"description\\\": \\\"Sipariş numarasına bağlı müşteri kimliğini, saklanan kişisel ver…"
}
```

A math row with reasoning (texts shortened):
```json
{
  "id": "000fa14c-a7b1-4608-937e-f25a9ff64f3c",
  "task_type": "Mathematical & Logical Reasoning",
  "task_subtype": "probability",
  "difficulty": "hard",
  "reasoning_effort": "low",
  "n_turns": 1,
  "has_tools": false,
  "messages": [
    {
      "role": "user",
      "content": "dolabımda 3 kırmızı, 4 mavi, 5 tane de siyah elbise var. gözlerimi kapayıp rastgele 2 tane seçtiğimde, bu ikisinin de aynı renk olma olasılığı kaçtır? hesaplaya…"
    },
    {
      "role": "assistant",
      "content": "Seçilen iki elbisenin aynı renk olma olasılığı 19/66'dır, yaklaşık %28.8.",
      "reasoning": "Önce toplam elbise sayısını buluyorum: 3+4+5=12. Rastgele 2 elbise seçmek için tüm olası kombinasyonlar C(12,2) = (12*11)/2 = 66. Aynı renkli kombinasyonları te…"
    }
  ]
}
```

## Training with it
Reasoning is kept separate from the answer, so it fits any chat template. For Qwen3-style templates, pass it as
`reasoning_content`; to train a model without thinking, drop it.
```python
import json
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
def to_chat(row, keep_reasoning=True):
    msgs = []
    for m in row["messages"]:
        d = {"role": m["role"], "content": m["content"]}
        if m["role"] == "assistant":
            if keep_reasoning and m["reasoning"]: d["reasoning_content"] = m["reasoning"]
            if m["tool_calls"]:
                d["tool_calls"] = [{"type": "function", "function": {"name": c["name"], "arguments": json.loads(c["arguments"])}}
                                   for c in m["tool_calls"]]
        if m["role"] == "tool": d["name"] = m["name"]
        msgs.append(d)
    return tok.apply_chat_template(msgs, tools=json.loads(row["tools"]) if row["tools"] else None, tokenize=False)
```
In multi-turn rows, the reasoning of earlier turns is usually dropped by chat templates, which keep only the last
turn's thinking. To learn every turn's reasoning, split a conversation into one sample per assistant turn.

## Contents and statistics
| config | task type | rows | multi-turn | with reasoning | user words (median) | answer words (median) | reasoning words (median) |
|---|---|---|---|---|---|---|---|
| `extractive_qa` | Extractive QA | 2,953 | 627 | 2,023 | 103 | 4 | 36 |
| `grammatical_error_correction` | Grammatical Error Correction (GEC) | 2,576 | 0 | 2,115 | 50 | 31 | 46 |
| `multiple_choice_qa` | MCQA | 2,330 | 0 | 1,890 | 74 | 1 | 62 |
| `summarization` | Summarization | 2,219 | 474 | 1,708 | 116 | 36 | 58 |
| `open_qa` | Open-ended Question Answering | 1,843 | 689 | 1,400 | 66 | 81 | 72 |
| `natural_language_inference` | NLI | 1,807 | 0 | 1,540 | 73 | 1 | 55 |
| `machine_translation` | Machine Translation | 1,660 | 0 | 918 | 57 | 47 | 51 |
| `math_and_logic` | Mathematical & Logical Reasoning | 1,501 | 579 | 1,292 | 84 | 58 | 119 |
| `text_classification` | Text Classification | 1,485 | 0 | 1,027 | 76 | 1 | 39 |
| `instruction_following` | Instruction Following | 1,311 | 792 | 972 | 95 | 83 | 95 |
| `code` | Code Generation & Understanding | 988 | 384 | 909 | 98 | 100 | 113 |
| `tool_calling` | Tool Use & Function Calling | 921 | 431 | 475 | 41 | 37 | 51 |
| `information_extraction` | Information Extraction | 850 | 166 | 706 | 92 | 19 | 68 |
| `casual_conversation` | Casual Conversation | 814 | 500 | 207 | 49 | 60 | 51 |
| `creative_writing` | Creative Writing | 729 | 307 | 402 | 56 | 108 | 52 |
| `planning` | Problem Solving & Planning | 556 | 211 | 472 | 94 | 137 | 116 |
| | **total** | **24,543** | **5,160** | **18,056** | 79 | 26 | 57 |

- **Reasoning effort:** none 6,487 (26.4%) · low 6,971 (28.4%) · medium 8,894 (36.2%) · high 2,026 (8.3%) · xhigh 165 (0.7%)
- **Reasoning mode:** effort 13,364 (54.5%) · auto 8,185 (33.3%) · off 2,994 (12.2%)
- **Difficulty:** medium 12,076 (49.2%) · easy 6,655 (27.1%) · hard 5,811 (23.7%) · expert 1 (0.0%)
- **User turns per conversation:** 1: 19,383 · 2: 3,309 · 3: 1,420 · 4: 323 · 5+: 108
- **Language mode:** tr_only 16,864 (68.7%) · tr_en 6,637 (27.0%) · en_source 1,042 (4.2%)
- **Culture:** everyday_tr 10,224 (41.7%) · general 8,249 (33.6%) · turkish_history 2,207 (9.0%) · proverb_idiom 2,128 (8.7%) · islamic_history 1,735 (7.1%)
- **User prompt style:** informal 6,958 (28.4%) · messy_typos 5,011 (20.4%) · exam 3,714 (15.1%) · conversational 3,267 (13.3%) · terse 2,901 (11.8%) · clean 1,786 (7.3%) · polite 874 (3.6%) · null 32 (0.1%)
- **Answerability:** answerable 23,440 (95.5%) · unanswerable 1,103 (4.5%)
- **Tool calling:** 921 rows, 1,307 calls; 419 rows make 2+ calls.
  - Subtypes: single_call 234, sequential_calls 179, no_tool_needed 153, parallel_calls 133, missing_argument 129, tool_error 93.
  - What they cover: a single call, independent calls made together, calls that need an earlier result, requests no
    declared tool serves, a missing required argument (asked for, then called), and a failing tool (reported honestly).
- **Answer length:** median 26 words, 90th percentile 127. Labels and spans are short; plans, code and stories are long.

## How it was made
- **Recipes:** every conversation starts from a sampled recipe (task, difficulty, user style, culture, number of
  turns, reasoning mode, topic) and is written in full by our own generator.
- **Reasoning:** sized to what each turn needs: none for labels and greetings, the most for multi-step math and code.
- **Quality:** every row passed automatic checks and a manual review before release.

## License and terms
Released under the **Apache License 2.0** (see `LICENSE`): use, modify and redistribute the data for any purpose,
including commercial use and model training, keeping the licence and its notices. The conversations were
generated with third-party large language model services.

## Citation
```bibtex
@misc{turkishchatmax25k2026,
  title     = {Turkish Chat Max 25K},
  author    = {cisimcik},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/cisimcik/turkish-chat-max-25k}
}
```