Update README.md

README.md

@@ -29,16 +29,18 @@ tags:
 - reasoning
 - llm
 - dataset
+- chatml
+- conversational
 - planning-assistant
 ---
 
 # Turkish Planning SFT
 
-**Turkish Planning SFT** is a large-scale synthetic instruction-following dataset designed to improve **planning capabilities** in Turkish Large Language Models (LLMs).
+**Turkish Planning SFT** is a large-scale synthetic instruction-following dataset designed to improve the planning capabilities of Turkish Large Language Models (LLMs).
 
-Instead of focusing on factual question answering, the dataset teaches models how to transform user goals into structured, practical, and actionable plans.
+Rather than focusing on factual question answering, the dataset teaches models how to transform user goals, requirements, and constraints into structured, practical, and actionable plans.
 
-The dataset is intended for **Supervised Fine-Tuning (SFT)** of Turkish instruction models.
+The dataset is intended for **Supervised Fine-Tuning (SFT)** and follows a conversation-oriented format compatible with modern chat models.
 
 ---
 
@@ -46,18 +48,19 @@ The dataset is intended for **Supervised Fine-Tuning (SFT)** of Turkish instruct
 
 - 🇹🇷 Entirely in Turkish
 - 🤖 Synthetic instruction-following dataset
-- 📋 Planning-oriented responses
+- 📋 Planning-oriented conversations
 - 🎯 Goal-driven tasks
 - 🧩 Diverse real-world scenarios
 - 📚 Multi-domain coverage
-- 🛣️ Long-form structured outputs
-- ⚙️ Designed for LLM Supervised Fine-Tuning (SFT)
+- 🛣️ Long-form structured responses
+- 💬 Chat-based data format
+- ⚙️ Compatible with modern SFT pipelines
 
 ---
 
 # Domains
 
-The dataset covers a broad range of planning scenarios, including but not limited to:
+The dataset covers a broad range of planning scenarios, including:
 
 - Business
 - Entrepreneurship
@@ -76,110 +79,84 @@ The dataset covers a broad range of planning scenarios, including but not limite
 - Research
 - Operations
 - Customer Service
-- Non-profit Organizations
 - Manufacturing
 - Agriculture
+- Non-profit Organizations
 
 ---
 
 # Dataset Format
 
-Each example follows the structure below.
+Each sample is represented as a conversation.
 
 ```json
-{
-  "messages": [
-    {
-      "role": "user",
-      "content": "..."
-    },
-    {
-      "role": "assistant",
-      "content": "..."
-    }
-  ],
-  "thinking": "...",
-  "_meta": {
-    "katman1_name": "...",
-    "katman2_name": "...",
-    "katman3_id": 12345
+[
+  {
+    "role": "user",
+    "content": "...",
+    "thinking": null,
+    "images": null,
+    "tool_calls": null
+  },
+  {
+    "role": "assistant",
+    "content": "...",
+    "thinking": "...",
+    "images": null,
+    "tool_calls": null
   }
-}
+]
 ```
 
 ---
 
 # Fields
 
-## messages
-
-Contains the user instruction and the assistant response.
-
-The assistant generates structured planning outputs such as:
-
-- Roadmaps
-- Timelines
-- Checklists
-- Milestone Plans
-- Weekly Plans
-- Monthly Plans
-- Multi-phase Strategies
-- Action Plans
-- Execution Plans
+| Field | Description |
+|--------|-------------|
+| `role` | Speaker role (`user` or `assistant`) |
+| `content` | User request or assistant response |
+| `thinking` | High-level planning strategy used during data generation. **This is not chain-of-thought.** |
+| `images` | Reserved for multimodal compatibility. Currently `null`. |
+| `tool_calls` | Reserved for tool-calling compatibility. Currently `null`. |
 
 ---
 
-## thinking
+# About the `thinking` Field
 
-The **thinking** field is **NOT** chain-of-thought.
+The **thinking** field is **NOT** chain-of-thought or hidden reasoning.
 
-Instead, it contains a high-level planning abstraction describing:
+Instead, it contains a high-level planning abstraction describing aspects such as:
 
-- Request analysis
-- Objectives
-- Constraints
-- Potential risks
-- Planning methodology
-- Response organization
-- Communication style
+- Understanding the user's objective
+- Identifying goals
+- Recognizing constraints
+- Considering potential risks
+- Selecting an appropriate planning strategy
+- Organizing the response
+- Choosing an appropriate communication style
 
-It does **not** contain hidden reasoning or intermediate reasoning traces.
-
----
-
-## _meta
-
-Metadata used during dataset generation.
-
-Example:
-
-```json
-{
-  "katman1_name": "Business",
-  "katman2_name": "Startup",
-  "katman3_id": 1204
-}
-```
-
-These fields are intended for categorization and filtering.
+It does **not** expose internal reasoning processes or intermediate inference steps.
 
 ---
 
 # Response Styles
 
-Examples may contain one or more planning formats including:
+Assistant responses include structured planning formats such as:
 
-- Step-by-step plans
-- Checklists
 - Roadmaps
 - Timelines
-- Priority lists
-- Milestone planning
-- Goal decomposition
-- Strategic planning
-- Learning plans
-- Improvement plans
-- Implementation strategies
+- Checklists
+- Action Plans
+- Execution Plans
+- Milestone Plans
+- Weekly Plans
+- Monthly Plans
+- Strategic Plans
+- Learning Plans
+- Goal Decomposition
+- Improvement Plans
+- Implementation Strategies
 
 ---
 
@@ -190,42 +167,43 @@ This dataset is suitable for:
 - Supervised Fine-Tuning (SFT)
 - Turkish Instruction Tuning
 - Planning Assistants
-- Productivity Assistants
 - Project Management Assistants
+- Productivity Assistants
 - Educational Assistants
 - Goal-Oriented AI Systems
-- Research on Planning Capabilities in LLMs
+- Research on Planning Capabilities in Turkish LLMs
 
 ---
 
 # Generation
 
-The dataset was generated using a hierarchical scenario generation pipeline.
+The dataset was synthetically generated using a hierarchical scenario generation pipeline.
 
 The generation process combines:
 
 - Hierarchical topic selection
-- Diverse personas
+- Diverse planning domains
+- Multiple personas
 - Variable constraints
-- Planning-oriented prompts
-- Structured response templates
+- Diverse planning templates
+- Structured conversational formatting
 
-This approach increases diversity while maintaining a consistent planning objective.
+The objective is to maximize planning diversity while maintaining consistent instruction-following behavior.
 
 ---
 
 # Limitations
 
 - Entirely synthetic.
-- Does not aim to be a factual knowledge dataset.
-- Focused on planning rather than domain expertise.
-- Responses should not replace professional advice in medical, legal, or financial contexts.
+- Designed specifically for planning and instruction-following tasks.
+- Does not aim to serve as a factual knowledge benchmark.
+- Responses should not replace professional advice in areas such as medicine, law, or finance.
 
 ---
 
 # Citation
 
-If you use this dataset in your work, please cite:
+If you use this dataset in your research, please cite:
 
 ```bibtex
 @dataset{colak2026turkishplanningsft,