Sable

An in-development personal project written in Rust designed to play Chess better than i can!

Included Features

• Sable uses a principal variation search plus quiescence search with iterative deepening
  • transposition table
  • move ordering
    • transposition table move ordering
    • previous PV move ordering
    • promotion ordering
    • SEE move ordering
  • null move pruning
  • late move reductions
  • history
    • quiet history
    • capture history
    • continuation history
    • counter moves
  • killers (2 per ply)
  • repetition handling
  • mate distance pruning
  • check extensions
  • singular extensions
  • internal iterative reductions
  • futility pruning
  • reverse futility pruning
  • late quiet pruning
  • capture SEE pruning
  • quiet SEE pruning
  • razoring
  • aspiration windows
  • correction history
    • pawn
    • minor
    • non-pawn
    • continuation
• An efficiently updatable neural network - Shard 3.2
  • (768x16hm->768)x2->8 arch
  • Trained on ~1.9 billion positions of selfplay data, 100% of which was DFRC games.
  • Trained using Bullet
• Lazy SMP for efficient multi-thread usage
• UCI protocol

Future Features

• A net with more complex input features and multi hidden layers
• Multilayer
• More layers after the first layer in my net
• A more complicated net with things like multiple layers and output buckets
• more elo

UCI Options

Name	Type	Default	Min/Max or Vars	Description
Hash	spin	16	1 / 32768	Transposition table size in MiB.
Threads	spin	1	1 / 256	Number of search threads.
Ponder	check	false		go ponder is held until ponderhit or stop.
MultiPV	spin	1	1 / 256	Number of principal variations to search and report.
UseSoftNodes	check	false		Treats go nodes as a soft node limit for datagen.
UCI_Chess960	check	false		Enables Chess960 FEN parsing and castling move notation.
UCI_ShowWDL	check	false		Adds WDL values to UCI info lines.
Move Overhead	spin	100	0 / 10000	Milliseconds reserved from time controls to avoid flagging.
Clear Hash	button			Clears the transposition table.
Eval File	string	embedded if compiled in, otherwise blank		Loads a native Shard NNUE file from disk, or embedded for the compiled-in net.

Strength

Version	My Estimate	CCRL 40/15	CCRL FRC 40/2	CCRL Blitz
3.0.0	3450	-	3551	-
2.0.0	3300	3258	3314	-
1.1.0	2900	2915	-	2932
1.0.0	2800	-	-	-

Project Details

My primary goal with Sable is to learn more about low-level programming and also training cool networks

With only a moderate proficiency in Rust, in-line completions were frequently used to assist with writing Rust syntax. While developing the code for NNUE, Mr GPT was consulted for explanations and snippets. LLM agents have modified the codebase for faster mass deletions or param changes.

You can create your own Sable build with cargo build --release.

Release Builds

The embedded Shard NNUE is read from data/quantised.bin when that file exists. Builds without an embedded net still compile, but search requires Eval File to be loaded before use.

Thanks :D

• Many members in the Stockfish discord server for their help with my questions (no particular order)
  • Chef
  • Dr Extension
  • Matt
  • Ciekce
  • DarkNeutrino
  • jb1729
  • Dan
  • Many others..!!
• Close friends who helped with datagen for my net
  • wnnb3dgy
  • HipHop
  • Tosiakowa
  • Bedthyme
• Members of MattBench for compute stuff