Upload MIMIC character checkpoints + model card

README.md

@@ -34,10 +34,7 @@ outputs controller inputs (main stick, c-stick, shoulder, buttons) at 60 Hz.
 
 | Character | Games | Val btn F1 | Val main F1 | Val loss | Step |
 |---|---|---|---|---|---|
-| **Fox** | 17,319 | 87.7% | ~55% | 0.77 | 31,065 |
-| **Falco** | 9,110 | 88.2% | 58.5% | 0.68 | 28,122 |
-| **Captain Falcon** | 9,404 | 89.9% | 52.2% | 0.71 | 26,814 |
-| **Luigi** | 1,951 | ~91% | ~60% | ~1.0 | 5,242 |
+| **Fox** | 17,319 | 87.1% | ~55% | 0.77 | 55,692 |
 
 ## Repo layout
 

fox/config.json

@@ -3,14 +3,14 @@
   "nhead": 8,
   "num_layers": 6,
   "dim_feedforward": 2048,
-  "dropout": 0.1,
-  "encoder_type": "hal_flat",
+  "dropout": 0.2,
+  "encoder_type": "mimic_flat",
   "d_intra": 256,
   "encoder_nlayers": 2,
   "k_query": 1,
   "scaled_emb": false,
   "max_seq_len": 256,
-  "pos_enc": "rope",
+  "pos_enc": "relpos",
   "rope_theta": 10000.0,
   "rope_learnable_freqs": false,
   "xpos_scale_base": 0.0,
@@ -26,11 +26,14 @@
   "n_stick_clusters": 37,
   "n_shoulder_bins": 3,
   "autoregressive_heads": true,
-  "hal_mode": true,
+  "mimic_mode": true,
   "lean_features": false,
+  "mimic_minimal_features": true,
+  "mimic_controller_encoding": true,
+  "n_controller_combos": 7,
+  "hal_mode": true,
   "hal_minimal_features": true,
   "hal_controller_encoding": true,
-  "n_controller_combos": 7,
   "num_stages": 6,
   "num_ports": 4,
   "num_characters": 27,
@@ -39,6 +42,6 @@
   "num_c_dirs": 9,
   "num_proj_types": 103,
   "num_proj_subtypes": 40,
-  "model_preset": "hal-rope",
-  "run_name": "fox-rope-v2"
+  "model_preset": "mimic",
+  "run_name": "fox-20260414-relpos"
 }

fox/metadata.json

@@ -1,16 +1,16 @@
 {
   "character": "Fox",
   "melee_enum": "FOX",
-  "run_name": "fox-rope-v2",
-  "global_step": 31065,
+  "run_name": "fox-20260414-relpos",
+  "global_step": 55692,
   "games_trained": 17319,
-  "val_btn_f1": "87.7%",
+  "val_btn_f1": "87.1%",
   "val_main_f1": "~55%",
   "val_loss": "0.77",
-  "n_params": 19591909,
+  "n_params": 26276389,
   "n_controller_combos": 7,
-  "model_preset": "hal-rope",
-  "pos_enc": "rope",
+  "model_preset": "mimic",
+  "pos_enc": "relpos",
   "no_self_inputs": false,
-  "training_notes": "Trained 2026-04-13, fox-rope-v2 run. v2 shards, RoPE position encoding (hal-rope preset), --self-inputs, dropout 0.1, 32K steps at batch 512. Replaces the legacy hal-7class-v2-long checkpoint which was trained without --self-inputs and stuck at val loss 2.27. The self-inputs flag was the real fix \u2014 RoPE vs relpos is secondary. Metrics land alongside Falco/CptFalcon/Luigi."
+  "training_notes": "Trained 2026-04-14, fox-20260414-relpos run. v2 shards, Shaw relative-position attention (mimic preset), --self-inputs, dropout 0.2, 65K steps at batch 512 (2x the standard schedule). Replaces the yesterday's RoPE Fox, which matched this on val metrics but relpos won the 2026-04-14 sweep on Falco so we standardized on it for all new runs. Best checkpoint is the step 55692 snapshot, closest to the lowest measured val loss."
 }

fox/model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:41ee9f42c3f5c8d03dfe799e2e8b08a142a88fe3328e295e99ebcf855162def0
-size 235274570
+oid sha256:03c11167493e4dbf11c8871135e82194714edde9ca522fd656cd008e9e74436d
+size 265142496