Model trained on chess "narratives" created from PGN notation from a large set of games downloaded from The Week in Chess (https://theweekinchess.com/). A script was run to convert the PGN notation to english text, and the model was finetuned on that. The approach is described in the paper Navigating Human Language Models with Synthetic Agents.

Useful Prompts:

• "The game begins"
• "In move X" // X can be a number between 1 and approximately 100
• "White/Black moves X from Y" // X is the piece (pawn, bishop, knight, rook, queen, king) and Y is the square (e.g. e2)
• "The game begins as white uses the X opening" // X is a known opening move such as Sicilian
• "White moves X from" // X is the piece (pawn, bishop, knight, rook, queen, king)
• "Black moves X from" // X is the piece (pawn, bishop, knight, rook, queen, king)

Citation:

@misc{feldman2020navigating,
      title={Navigating Human Language Models with Synthetic Agents}, 
      author={Philip Feldman and Antonio Bucchiarone},
      year={2020},
      eprint={2008.04162},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}