suman-kalavagunta/chess-coach-turningpoints

Chess Coach – Turning Point Explanations Dataset

Overview

This repository contains a curated, engine-grounded dataset for training language models to explain chess mistakes and turning points in a human coaching style.

The goal is explainability and pedagogy, not move calculation or engine strength.

---

What this dataset is (and is not)

✅ This dataset is for

• Training LLMs to explain evaluation swings
• Teaching coaching tone, structure, and pedagogy
• Demonstrating engine-grounded LLM pipelines
• Research in explainable AI for games

❌ This dataset is NOT for

• Training a chess engine
• Improving raw move strength
• Replacing Stockfish or other engines
• Competitive or tournament play

Important:
The language model does not calculate chess.
All chess truth comes from Stockfish and deterministic rules.

---

Dataset Structure

Each example corresponds to a single turning point extracted from a real game.

Format: JSONL (chat-style)

{
  "messages": [
    {
      "role": "user",
      "content": "<structured prompt with engine evals and features>"
    },
    {
      "role": "assistant",
      "content": "<coach-style explanation>"
    }
  ]
}

---

What the prompt contains

Each prompt typically includes:

• Engine evaluation before the move
• Engine evaluation after the move
• Net centipawn swing
• Best engine continuation
• Deterministic features such as:
  • Development gap
  • King safety signals
  • Tactical motifs
  • Material imbalance
• Output constraints (format, tone, tense)

---

What the response contains

Each response follows a coaching template, usually including:

1. Net result summary
2. What changed in the position
3. What to do differently next time
4. Reflective questions
5. A short practice drill

---

Dataset Splits

The dataset follows standard Hugging Face conventions:

• train – ~90% of examples
• validation – ~10% of examples

Suitable for Supervised Fine-Tuning (SFT) using LoRA / QLoRA.

---

How the dataset was generated

High-level pipeline:

1. Parse multi-game PGN files
2. Identify candidate turning points
3. Evaluate positions using Stockfish
4. Extract deterministic chess features in Python
5. Construct structured prompts
6. Generate explanations using a teacher LLM
7. Apply strict validation rules
8. Emit clean (prompt → response) pairs

This design intentionally separates:

• Truth → engines + rules
• Explanation → language model

---

Intended Use

This dataset is well-suited for:

• Fine-tuning instruction/chat LLMs
• Building chess coaching or educational tools
• Studying small, high-quality datasets for behavior shaping
• Demonstrating hybrid symbolic + neural systems

---

Limitations

• Explanations reflect engine evaluations, not human intuition
• Focused on turning points, not full game narration
• Style reflects the constraints used during generation
• Coverage depends on the source PGN data

---

License

This dataset is released under the Apache 2.0 License.

---

Acknowledgements

• Stockfish – for world-class chess evaluation
• Hugging Face – for open tooling and hosting
• The open chess community – for publicly available PGN data

---

Disclaimer

This dataset and any derived models are provided for educational and research purposes only.