| { | |
| "architecture": "ChessTransformer200M", | |
| "params": "~204M", | |
| "vocab_version": "compact", | |
| "config": { | |
| "encoder_dim": 256, | |
| "encoder_type": "fused", | |
| "encoder_conv_blocks": 2, | |
| "normalize_stm": false, | |
| "hidden_dim": 1024, | |
| "num_layers": 16, | |
| "num_heads": 16, | |
| "ffn_ratio": 4, | |
| "dropout": 0.1, | |
| "policy_head_dim": 512, | |
| "value_hidden": 512, | |
| "use_pos_embed": true, | |
| "n_ctx_tokens": 4, | |
| "value_head_type": "cls", | |
| "n_value_classes": 3, | |
| "use_swiglu": false, | |
| "use_rel_bias": false, | |
| "gradient_checkpointing": false, | |
| "full_dim_attention": false | |
| }, | |
| "training": { | |
| "experiment": "exp189_200m_maxelo_policy", | |
| "base_model": "avewright/chess-transformer-200m-compact-soft", | |
| "soft_data": "avewright/exp186-sf-multipv-2m + avewright/exp190-phase-deep-soft", | |
| "latest_step": 2851, | |
| "best_metric_blended_top1": 0.3965625, | |
| "inference": "policy argmax (no MCTS)" | |
| }, | |
| "elo_eval": { | |
| "estimated_elo": 1700, | |
| "bracket": [ | |
| 1500, | |
| 1800 | |
| ], | |
| "protocol": "elo_eval_latest.py vs SF18 UCI_LimitStrength, openings x colors, syzygy+book" | |
| } | |
| } |