pkrisz/renegade-training-data

Training data for Renegade (chess engine)

Here are the positions that the neural networks of the Renegade chess engine were trained on. These neural networks are necessary for accurately evaluating positions, and significantly contribute to the playing strength. As the engine progresses, the datasets are regenerated and replaced over time.

The data is original, and were generated using the internal data generation utility included in various development versions of the chess engine. During this process, the engine plays millions of games against itself from randomized positions. The moves played are determined via low-depth searches. This process is repeated, and the engine should become stronger with each iteration.

Note: These datasets are released into the public domain, and can be freely used for research, testing, and training.
However, when writing a competitive chess engine, directly training on datasets created by others is frowned upon by a significant part of the community.

Format

These datasets are stored in viriformat, with each file containing millions of games. It is a lossless binary format, the games are perfectly reconstructable. Community-made tools are available to further process these.

The games are stored as:

• 32 bytes for the starting position
• 4 bytes per move for the best move and evaluation
• 4 bytes of zeros to signal the end of a game

Files & statistics

File	Renegade version	Version ID	Soft node limit	Number of games (millions)	Total position count (billions)	Mean game length	DFRC
251202.vf	1.2.??	1	10k	14.1	1.814	128.2
260112.vf	1.2.13	2	10k	11.7	1.480	126.4
dfrc260126.vf	1.2.25	3	20k	10.7	1.375	129.1	yes
260217.vf	1.2.32	4	10k	23.6	3.003	127.2
260312.vf	1.2.40	5	10k	30.8	3.830	124.5