File size: 8,516 Bytes
cb93d01 a6e27d6 cb93d01 a6e27d6 cb93d01 a6e27d6 cb93d01 a6e27d6 09be58c a6e27d6 09be58c a6e27d6 cb93d01 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 | {
"description": "Checkpoints for 'Temporal Conditioning for Human Behavioral Cloning in Chess' (IEEE GEM 2026).",
"note": "Parameter counts are unique parameters. A naive sum over state_dict() is 4,194,304 larger because the GAB weight is shared across all 8 blocks and counted 8 times.",
"checkpoints": [
{
"file": "marvin_23m_cond_1M.pt",
"role": "23M time-conditioned, 1M steps. Headline arm of the controlled ablation (Tables I-III, Fig. 1).",
"source_filename": "chessformer_maia3-23m_elo-endpoints_time-tokens_tempotok_s67_step1000000.pt",
"size_bytes": 280134073,
"steps": 1000000,
"seed": 67,
"data_seed": 42,
"parameters": 23315934,
"time_conditioned": true,
"config_dict": {
"history": 8,
"num_blocks": 8,
"mlp_ratio": 2.0,
"dim_emb": 128,
"dropout": 0.0,
"use_gab": true,
"use_rms_norm": true,
"omit_qkv_biases": true,
"activation": "gelu",
"time_encoder": "tokens",
"time_token_encoder": "interp",
"time_featurize": "log1p",
"per_position_clocks": false,
"shrink_global_time": false,
"per_position_time_deltas": false,
"no_time_context": false,
"num_piece_types": 13,
"gradient_checkpointing": false,
"dim_vit": 512,
"head_hid_dim": 512,
"num_heads": 16,
"gab_gen_size": 128,
"gab_per_square_dim": 32,
"gab_intermediate_dim": 128,
"tempo_history_token": true,
"zero_time_history": false,
"elo_encoder": "endpoints"
},
"train_command": "python train.py --config 23m --batch-size 256 --grad-accum-steps 4 --lr 1.25e-4 --seed 67 --data-seed 42 --data-dir data/train --val-data-dir data/val",
"sha256": "ee82bd936f492bb5cce97baf4335172bb7f6dca71a6652d7eedae0ed7bfcc57f"
},
{
"file": "marvin_23m_blind_1M.pt",
"role": "23M time-blind twin, 1M steps. Identical backbone, data and schedule; receives zeroed temporal inputs. Training stopped here, so this is also the latest time-blind checkpoint.",
"source_filename": "chessformer_23m_elo-endpoints_time-blind_s67_step1000000.pt",
"size_bytes": 279261361,
"steps": 1000000,
"seed": 67,
"data_seed": 42,
"parameters": 23249117,
"time_conditioned": false,
"config_dict": {
"history": 8,
"num_blocks": 8,
"mlp_ratio": 2.0,
"dim_emb": 128,
"dropout": 0.0,
"use_gab": true,
"use_rms_norm": true,
"omit_qkv_biases": true,
"activation": "gelu",
"gab_additive_bias": true,
"gab_qk_rotation": false,
"time_encoder": "tokens",
"time_token_encoder": "interp",
"time_featurize": "log1p",
"tempo_history_token": true,
"per_position_clocks": false,
"shrink_global_time": false,
"per_position_time_deltas": false,
"no_time_context": false,
"num_piece_types": 13,
"gradient_checkpointing": false,
"dim_vit": 512,
"head_hid_dim": 512,
"num_heads": 16,
"gab_gen_size": 128,
"gab_per_square_dim": 32,
"gab_intermediate_dim": 128,
"zero_time_history": false,
"no_time_conditioning": true,
"elo_encoder": "endpoints"
},
"train_command": "python train.py --config 23m --batch-size 256 --grad-accum-steps 4 --lr 1.25e-4 --seed 67 --data-seed 42 --data-dir data/train --val-data-dir data/val --no-time-conditioning",
"sha256": "b8b6d40132bb1f21e12cdc801bd674efe07de106a6cb36cc6eeecf1073c63f54"
},
{
"file": "marvin_79m_cond_1M.pt",
"role": "79M time-conditioned at 1M steps, for step-matched comparison with the 23M pair.",
"source_filename": "chessformer_79m_elo-endpoints_time-tokens_tempotok_s67_step1000000.pt",
"size_bytes": 956000269,
"steps": 1000000,
"seed": 67,
"data_seed": 42,
"parameters": 79641566,
"time_conditioned": true,
"config_dict": {
"history": 8,
"num_blocks": 8,
"mlp_ratio": 2.0,
"dim_emb": 128,
"dropout": 0.0,
"use_gab": true,
"use_rms_norm": true,
"omit_qkv_biases": true,
"activation": "gelu",
"gab_additive_bias": true,
"gab_qk_rotation": false,
"time_encoder": "tokens",
"time_token_encoder": "interp",
"time_featurize": "log1p",
"tempo_history_token": true,
"per_position_clocks": false,
"shrink_global_time": false,
"per_position_time_deltas": false,
"no_time_context": false,
"num_piece_types": 13,
"gradient_checkpointing": false,
"dim_vit": 1024,
"head_hid_dim": 1024,
"num_heads": 32,
"gab_gen_size": 128,
"gab_per_square_dim": 32,
"gab_intermediate_dim": 128,
"zero_time_history": false,
"elo_encoder": "endpoints"
},
"train_command": "python train.py --config 79m --batch-size 256 --grad-accum-steps 4 --lr 1e-4 --seed 67 --data-seed 42 --data-dir data/train --val-data-dir data/val",
"sha256": "8bb522116f1e9edf4c20a4292216addb0983e2232d9a6ff37a248330c73aca84"
},
{
"file": "marvin_23m_cond_1700k.pt",
"role": "23M time-conditioned, 1.7M steps. Longest-trained 23M and the strongest by validation loss.",
"source_filename": "chessformer_maia3-23m_elo-endpoints_time-tokens_tempotok_s67_latest.pt",
"size_bytes": 280093471,
"steps": 1700000,
"seed": 67,
"data_seed": 42,
"parameters": 23315934,
"time_conditioned": true,
"config_dict": {
"history": 8,
"num_blocks": 8,
"mlp_ratio": 2.0,
"dim_emb": 128,
"dropout": 0.0,
"use_gab": true,
"use_rms_norm": true,
"omit_qkv_biases": true,
"activation": "gelu",
"time_encoder": "tokens",
"time_token_encoder": "interp",
"time_featurize": "log1p",
"per_position_clocks": false,
"shrink_global_time": false,
"per_position_time_deltas": false,
"no_time_context": false,
"num_piece_types": 13,
"gradient_checkpointing": false,
"dim_vit": 512,
"head_hid_dim": 512,
"num_heads": 16,
"gab_gen_size": 128,
"gab_per_square_dim": 32,
"gab_intermediate_dim": 128,
"tempo_history_token": true,
"zero_time_history": false,
"elo_encoder": "endpoints"
},
"train_command": "python train.py --config 23m --batch-size 256 --grad-accum-steps 4 --lr 1.25e-4 --seed 67 --data-seed 42 --data-dir data/train --val-data-dir data/val",
"sha256": "10679ea7b74e0d924a31c65977f2b2563a305ca3ba159d891cd0251252698861"
},
{
"file": "marvin_79m_cond_1650k.pt",
"role": "79M time-conditioned, 1.65M steps. The 79M model the paper evaluates against released Maia-3 (Fig. 2, Table IV).",
"source_filename": "chessformer_79m_elo-endpoints_time-tokens_tempotok_s67_latest.pt",
"size_bytes": 955983347,
"steps": 1650000,
"seed": 67,
"data_seed": 42,
"parameters": 79641566,
"time_conditioned": true,
"config_dict": {
"history": 8,
"num_blocks": 8,
"mlp_ratio": 2.0,
"dim_emb": 128,
"dropout": 0.0,
"use_gab": true,
"use_rms_norm": true,
"omit_qkv_biases": true,
"activation": "gelu",
"gab_additive_bias": true,
"gab_qk_rotation": false,
"time_encoder": "tokens",
"time_token_encoder": "interp",
"time_featurize": "log1p",
"tempo_history_token": true,
"per_position_clocks": false,
"shrink_global_time": false,
"per_position_time_deltas": false,
"no_time_context": false,
"num_piece_types": 13,
"gradient_checkpointing": false,
"dim_vit": 1024,
"head_hid_dim": 1024,
"num_heads": 32,
"gab_gen_size": 128,
"gab_per_square_dim": 32,
"gab_intermediate_dim": 128,
"zero_time_history": false,
"elo_encoder": "endpoints"
},
"train_command": "python train.py --config 79m --batch-size 256 --grad-accum-steps 4 --lr 1e-4 --seed 67 --data-seed 42 --data-dir data/train --val-data-dir data/val",
"sha256": "047e914d0d71be607b95c2a93d2e255ab3c97ff1b7c63602f3c1064a00709628"
}
]
} |