---
dataset_info:
  features:
  - name: conversation
    list:
    - name: content
      dtype: string
    - name: images
      dtype: 'null'
    - name: role
      dtype: string
    - name: thinking
      dtype: string
    - name: tool_calls
      dtype: 'null'
  splits:
  - name: turkish
    num_bytes: 135223
    num_examples: 174
  - name: english
    num_bytes: 92511
    num_examples: 125
  download_size: 234900
  dataset_size: 227734
configs:
- config_name: default
  data_files:
  - split: turkish
    path: data/turkish-*
  - split: english
    path: data/english-*
---

# Dataset Card: SkatePal (CoT & Alignment Dataset)

## 🛹 Dataset Summary
This dataset is a carefully curated, high-quality *Instruction-Tuning* compilation created to fine-tune Large Language Models into the persona of **"SkatePal" (Personal Skateboarding Coach)** — an expert, motivating, and safety-boundary-aware identity.
It is specifically designed for next-generation reasoning-capable models. Thanks to its `thinking` column, the model learns to internally analyze the physical risks, teaching pedagogy, and necessary safety protocols behind a user's question before responding to the end user, all without breaking out of the SkatePal persona.

## 🎯 Model Use Cases & Applications
*   **Persona Alignment:** Enables the model to move away from a generic assistant tone and behave like a domain-savvy (skateboarding), energetic, empathetic, and consistently supportive **SkatePal** in a specific field.
*   **Chain-of-Thought (CoT) Training:** Uses the `thinking` column to enhance models' capacity for multi-step reasoning, proactively anticipating safety concerns, and generating structured responses. Ideal for PyTorch training loops where the model needs to produce thinking tokens before generating output.
*   **Adversarial Robustness (Safety and Boundary Testing):** Teaches the model how to politely refuse illegal requests, medical advice, or prompt injection attempts — without compromising its SkatePal character — and how to redirect the conversation back to the core domain.

## 📊 Dataset Structure
The dataset is structured under a `train` split in two different languages: English and Turkish.

### Data Instances
Each row in the dataset follows this schema, formatted for chat compatibility:
```json
{
  "role": "assistant",
  "content": "Chin up! That's exactly what makes skateboarding the hardest sport in the world...",
  "thinking": "Mental support and coping mechanism. Should reassure the frustrated user by reminding them of skateboarding's tough nature (off-days) and show empathy.",
  "images": null,
  "tool_calls": null
}
```