Dataset Card: Chess-Puzzles-RLVR

Dataset Summary

This dataset is a highly processed and stratified collection of approximately 5 million chess puzzles, ranging from Elo ratings of 400 to 3300. It is specifically designed for Curriculum Learning and Reinforcement Learning (RLVR) agents.

Unlike standard puzzle datasets, this version is pre-sorted and split into "rating buckets" to ensure that training, validation, and testing sets maintain an identical difficulty distribution across the entire spectrum.

---

Dataset Structure

Data Instances

Each instance represents a unique chess puzzle with a starting position (FEN) and the correct sequence of moves (UCI).

Key Fields (Schema)

• fen (string): The Forsyth-Edwards Notation representing the board state before the first move of the puzzle.
• uci_moves (list of strings): The sequence of best moves in Universal Chess Interface (UCI) format (e.g., ["e2e4", "e7e5"]).
• rating (int): The difficulty rating of the puzzle (Elo).
• tags (list of strings): Tactical motifs associated with the puzzle (e.g., ["fork", "sacrifice", "mateIn2"]).
• turn (string): Indicates which side is to move in the starting FEN ("White" or "Black").

Data Splits

The dataset is split into three parts, with each split containing a proportionate amount of data from every 100-point rating interval:

Split	Percentage	Purpose
Train	85%	Primary data for model training.
Validation	10%	Monitoring performance and preventing catastrophic forgetting across difficulty tiers.
Test	5%	Final holdout set for objective evaluation.

---

Creation Process

1. Data Cleaning and Transformation

The dataset was transformed from a raw chess puzzle format using the python-chess library. The following steps were taken for every row:

• Turn Extraction: The active color was parsed directly from the FEN string.
• String Tokenization: Raw space-separated strings for moves and tags were converted into clean Python lists for easier model consumption.
• Feature Pruning: Redundant boolean flags (e.g., white_kingside, board) were removed to reduce the dataset footprint and focus strictly on necessary state representation.

2. Stratified Bucketing

To facilitate curriculum learning, the dataset underwent a unique Stratified Bucketing process:

1. The entire dataset was sorted globally by rating.
2. The data was partitioned into 29 buckets, each representing a 100-point rating range (e.g., 400-500, 501-600, ..., 3200-3300).
3. The 85/10/5 split was applied locally within each bucket.
4. These local splits were then re-concatenated into the final global train, validation, and test splits.

This ensures that whether the model is training on "easy" or "hard" data, the validation set always provides a statistically accurate reflection of the model's ability across the entire difficulty spectrum.

---

Usage Considerations

This dataset is optimized for a Sliding Window Sampler. During training, it is recommended to:

1. Sample 80% of your batch from the model's current "target" rating bucket.
2. Sample 20% from all previously learned (easier) buckets to maintain tactical proficiency and prevent regression.