canbingol/turkish-conversational-tool-call

Dataset Limitation & Experimental Results: Based on our experiments, the conversations built around the tools in this dataset are overly synthetic and unnatural. As a result, fine-tuned models do not show any significant performance improvement, and in some benchmarks, their scores slightly degrade by 1–2 points.

Turkish Tool Calling Experiments Dataset

A curated dataset of 1,810 Turkish single-turn and multi-turn conversations designed for training and evaluating tool/function calling capabilities in Turkish LLMs.

Dataset Structure

The dataset is split into train (1,500 examples) and test (310 examples).

Columns

• tools (list): List of available tool definitions in standard JSON Schema / OpenAI format.
• messages (list): Multi-turn conversation history with standard roles (system, user, assistant, tool), including tool_calls and tool execution outputs.
• tool_call_type (string): The structural pattern of tool usage in the conversation:
  • single_call: Exactly one tool call in a single turn.
  • parallel_call: Multiple tools executed simultaneously in one turn.
  • sequential_call: Tool calls executed step-by-step across multiple turns.
  • multi_turn_parallel_call: Multi-turn interaction containing parallel tool executions.
  • no_call: Conversations that require no tool calls (out-of-scope requests, general chitchat, refusals).

with_think config

Train split only (1,128 conversations). Every assistant decision point (each message with tool_calls, and each text reply that directly follows a user message) has a short Turkish reasoning chain prepended to its content as <think>...</think>. No separate think tool is used, so the reasoning works with any tool list. Reasoning was generated with DeepSeek-V4.1-Flash given the tools, the conversation so far and the gold decision. 372 conversations whose gold tool calls contained arguments not grounded in the user message, a previous tool result or a tool default were dropped.

ds = load_dataset("canbingol/turkish-conversational-tool-call", "with_think", split="train")

Distribution

Call Type	Train (1,500)	Test (310)	Total (1,810)
parallel_call	543	129	672
sequential_call	375	64	439
multi_turn_parallel_call	272	48	320
single_call	249	56	305
no_call	61	13	74
Total	1,500	310	1,810

Quickstart

from datasets import load_dataset

# Load both splits
dataset = load_dataset("canbingol/turkish-tool-call-experiments")
print(dataset)

# Access train and test splits
train_ds = dataset["train"]
test_ds = dataset["test"]

print(f"Train size: {len(train_ds)}")
print(f"Test size: {len(test_ds)}")
print(train_ds[0])