Update puff checkpoint + metadata

puff/config.json

@@ -1,8 +1,8 @@
 {
-  "d_model": 768,
-  "nhead": 12,
-  "num_layers": 6,
-  "dim_feedforward": 3072,
+  "d_model": 1024,
+  "nhead": 16,
+  "num_layers": 12,
+  "dim_feedforward": 4096,
   "dropout": 0.1,
   "encoder_type": "mimic_flat",
   "d_intra": 256,
@@ -17,7 +17,8 @@
   "attn_variant": "standard",
   "n_kv_heads": 0,
   "use_rmsnorm": false,
-  "use_swiglu": true,
+  "use_swiglu": false,
+  "use_input_gate": false,
   "stick_loss": "clusters",
   "btn_loss": "focal",
   "no_opp_inputs": true,
@@ -28,11 +29,11 @@
   "autoregressive_heads": true,
   "mimic_mode": true,
   "lean_features": false,
-  "mimic_minimal_features": true,
+  "mimic_minimal_features": false,
   "mimic_controller_encoding": true,
   "n_controller_combos": 7,
   "hal_mode": true,
-  "hal_minimal_features": true,
+  "hal_minimal_features": false,
   "hal_controller_encoding": true,
   "num_stages": 6,
   "num_ports": 4,
@@ -42,6 +43,6 @@
   "num_c_dirs": 9,
   "num_proj_types": 103,
   "num_proj_subtypes": 40,
-  "model_preset": "mimic-xl",
-  "run_name": "puff-20260418-relpos-xl"
+  "model_preset": "mimic-xxl",
+  "run_name": "puff-20260419-mimic-xxl"
 }

puff/metadata.json

@@ -1,16 +1,16 @@
 {
   "character": "Jigglypuff",
   "melee_enum": "JIGGLYPUFF",
-  "run_name": "puff-20260418-relpos-xl",
-  "global_step": 29757,
-  "games_trained": 33585,
-  "val_btn_f1": "91.3%",
-  "val_main_f1": "51.7%",
-  "val_loss": "0.6766",
-  "n_params": 50590757,
+  "run_name": "puff-20260419-mimic-xxl",
+  "global_step": 30738,
+  "n_params": 166949413,
   "n_controller_combos": 7,
-  "model_preset": "mimic-xl",
+  "model_preset": "mimic-xxl",
   "pos_enc": "relpos",
-  "no_self_inputs": false,
-  "training_notes": "Trained 2026-04-18 on the full Puff HuggingFace set (33,327 .slp files \u2192 33,585 games after quality filters, 168 \u00d7 4GB shards). v2 shards, relpos attention, --self-inputs, 32,768 steps at eff_bs 512 in 83 min on 2\u00d7RTX 5090. Uses the `mimic-xl` preset: d_model=768, 6 layers, 12 heads, d_ff=3072, SwiGLU FFN, dropout 0.1 \u2014 ~44M params, a 2.2\u00d7 scaled-up variant of the standard `mimic` preset. Beat the matching standard-size puff baseline (val 0.6890) by 1.8% (run qeka6rq8 would be an apples-to-apples reference run for capacity comparisons)."
+  "val_loss": "0.6432",
+  "best_step": 30738,
+  "val_btn_f1": "91.4%",
+  "val_main_f1": "53.3%",
+  "val_shldr_f1": "91.0%",
+  "val_cdir_f1": "68.7%"
 }

puff/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:65fdc96b775bf0efd240fc76cc194fd4d8f3a8ec2a5e31f5cdc429a970ad2107
-size 556909050
+oid sha256:e0bcad2e1f0250c07a4574aa346e73556d30c71306a1ff76563e45c21ba920ad
+size 1902993716