README.md

---
license: mit
tags:
- chess
task_categories:
- reinforcement-learning
---
# ChessExplained_500k_v0 Dataset

## Overview
This dataset contains approximately 500,000 chess positions with explanations for optimal moves, prepared for Supervised Fine-Tuning (SFT). The data is tokenized using **Qwen/Qwen3-8B** tokenizer.

## Format
- **File**: JSONL format (one JSON object per line)
- **Fields**:
  - `text`: Full conversation in Qwen chat format with special tokens
  - `fen`: Position in FEN notation
  - `move`: Best move in UCI notation
  - `explanation`: Natural language explanation for the move

## Explanation Generation Process

Each explanation is generated through the following pipeline:

1. **Position Analysis**: Extract current position features from the FEN
2. **Best Move Calculation**: Determine optimal move using Stockfish engine
3. **Move Application**: Apply the move and extract new position features
4. **Position Evaluation**: Use Stockfish evaluation to classify the position:
   - Winning (|eval| > 2)
   - Drawish (|eval| ≤ 2)
5. **Feature Extraction**: Identify key tactical/strategic elements by comparing position features:
   - Attacking new pieces
   - Escaping attacks
   - Forced checkmates
   - Captures
   - Pawn promotions
   - Checks
   - Center control
   - Pins
   - King safety improvements
   - Rooks on 7th rank

## Explanation Template

Explanations are generated by comparing position features before and after the best move. The system detects tactical/strategic changes and constructs natural language explanations.

**Standard format:**
```
After {piece} {action} {square}, this causes {color} to {reason(s)}. 
So {move_SAN} is the most logical. {evaluation}.
```

**Detected features include:**
- Checkmate delivery
- Castling (kingside/queenside)
- Giving check
- Captures
- Pawn promotions
- Escaping attacks
- Attacking new pieces
- Gaining center control
- Creating pins
- Placing rook on 7th/2nd rank

**Fallback:** When no specific tactical features are detected, the explanation describes piece development (early game) or repositioning (mid/late game).

## Example Entry

```json
{
  "text": "<|im_start|>user\nYou are an expert chess player...",
  "fen": "rnbq1rk1/ppp1bpp1/4pn1p/3p4/2PP4/2N1PN2/PP1B1PPP/R2QKB1R b KQ - 0 7",
  "move": "c7c5",
  "explanation": "After Pawn moves to c5, this causes Black to attacks the pawn on d4. So c5 is the most logical. Position is drawish."
}
```

```
<|im_start|>user
You are an expert chess player looking at the following position in FEN format:
r3kb1N/pp2p2p/n5p1/2p5/2BP2q1/7b/PPP2P2/R1BQK3 b q - 3 16
Briefly, FEN describes chess pieces by single letters [PNBRKQ] for white and [pnbrkq] for black. The pieces found in each rank are specified, starting at the top of the board (a8..h8) and describing all eight ranks. Within each rank, all 8 positions must be specified, with one or more empty squares noted with a digit [1..8]. For example, /8/ is an empty rank (no pieces), while /4P3/ specifies four empty squares, a white pawn, and three more empty squares.

Here is an additional visualization of the board (♔♕♖♗♘♙ = White pieces, ♚♛♜♝♞♟ = Black pieces):

   a b c d e f g h  
  +---------------+
8 | ♜  ·  ·  ·  ♚  ♝  ·  ♘ | 8
7 | ♟  ♟  ·  ·  ♟  ·  ·  ♟ | 7
6 | ♞  ·  ·  ·  ·  ·  ♟  · | 6
5 | ·  ·  ♟  ·  ·  ·  ·  · | 5
4 | ·  ·  ♗  ♙  ·  ·  ♛  · | 4
3 | ·  ·  ·  ·  ·  ·  ·  ♝ | 3
2 | ♙  ♙  ♙  ·  ·  ♙  ·  · | 2
1 | ♖  ·  ♗  ♕  ♔  ·  ·  · | 1
  +---------------+
   a b c d e f g h  

The current side to move is black.
The possible legal moves for the side to move are: ['f8g7', 'f8h6', 'e8d8', 'e8d7', 'a8d8', 'a8c8', 'a8b8', 'a6b8', 'a6c7', 'a6b4', 'g4c8', 'g4d7', 'g4e6', 'g4h5', 'g4g5', 'g4f5', 'g4h4', 'g4f4', 'g4e4', 'g4d4', 'g4g3', 'g4f3', 'g4g2', 'g4e2', 'g4g1', 'g4d1', 'h3g2', 'h3f1', 'e8c8', 'c5d4', 'h7h6', 'e7e6', 'b7b6', 'g6g5', 'h7h5', 'e7e5', 'b7b5'].

Your task is to select the best move for the side to move. Please choose a move from this list, and output it in the following format:
<uci_move>one_move_from_the_list</uci_move>.
An example of a valid move for this position is <uci_move>f8g7</uci_move>.
Think about the move and output it in the following format:
<think>your_thinking</think>
<uci_move>one_move_from_the_list</uci_move>.
<|im_end|>
<|im_start|>assistant
<think>
After Queen moves to g1, this causes Black to give check, moves away from attack, attacks the pawn on f2, creates a pin. So Qg1+ is the most logical. Black has a huge advantage.
</think>

<uci_move>g4g1</uci_move><|im_end|>
```