---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: black_player_name
    dtype: string
  - name: white_player_name
    dtype: string
  - name: winner
    dtype: string
  - name: black_player_elo
    dtype: float64
  - name: white_player_elo
    dtype: float64
  - name: winner_elo
    dtype: float64
  splits:
  - name: train
    num_bytes: 114775572
    num_examples: 93970
  download_size: 80385458
  dataset_size: 114775572
---
# go_pgn_string_v2

It is parsed from https://dl.fbaipublicfiles.com/elfopengo/analysis/data/gogod_commentary_sgfs.gzip.  
It contains professional game of Go ever played (~100k games drawn from GoGoD), evaluated by Meta AI's [ELF OpenGo](https://ai.meta.com/research/publications/elf-opengo-an-analysis-and-open-reimplementation-of-alpha-zero/).  
The leftmost variation of the game tree is taken in SGF format and translate it into PGN like format.  
Due to autoregressive nature of decoder, a special token '>' is used to denote the move by the winner of the game.