Add Puff (Jigglypuff) checkpoint: mimic-xl (768d/6L/12h/SwiGLU) val 0.6766
Browse files- puff/config.json +7 -7
- puff/metadata.json +10 -10
- puff/model.pt +2 -2
puff/config.json
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
{
|
| 2 |
-
"d_model":
|
| 3 |
-
"nhead":
|
| 4 |
"num_layers": 6,
|
| 5 |
-
"dim_feedforward":
|
| 6 |
-
"dropout": 0.
|
| 7 |
"encoder_type": "mimic_flat",
|
| 8 |
"d_intra": 256,
|
| 9 |
"encoder_nlayers": 2,
|
|
@@ -17,7 +17,7 @@
|
|
| 17 |
"attn_variant": "standard",
|
| 18 |
"n_kv_heads": 0,
|
| 19 |
"use_rmsnorm": false,
|
| 20 |
-
"use_swiglu":
|
| 21 |
"stick_loss": "clusters",
|
| 22 |
"btn_loss": "focal",
|
| 23 |
"no_opp_inputs": true,
|
|
@@ -42,6 +42,6 @@
|
|
| 42 |
"num_c_dirs": 9,
|
| 43 |
"num_proj_types": 103,
|
| 44 |
"num_proj_subtypes": 40,
|
| 45 |
-
"model_preset": "mimic",
|
| 46 |
-
"run_name": "puff-
|
| 47 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"d_model": 768,
|
| 3 |
+
"nhead": 12,
|
| 4 |
"num_layers": 6,
|
| 5 |
+
"dim_feedforward": 3072,
|
| 6 |
+
"dropout": 0.1,
|
| 7 |
"encoder_type": "mimic_flat",
|
| 8 |
"d_intra": 256,
|
| 9 |
"encoder_nlayers": 2,
|
|
|
|
| 17 |
"attn_variant": "standard",
|
| 18 |
"n_kv_heads": 0,
|
| 19 |
"use_rmsnorm": false,
|
| 20 |
+
"use_swiglu": true,
|
| 21 |
"stick_loss": "clusters",
|
| 22 |
"btn_loss": "focal",
|
| 23 |
"no_opp_inputs": true,
|
|
|
|
| 42 |
"num_c_dirs": 9,
|
| 43 |
"num_proj_types": 103,
|
| 44 |
"num_proj_subtypes": 40,
|
| 45 |
+
"model_preset": "mimic-xl",
|
| 46 |
+
"run_name": "puff-20260418-relpos-xl"
|
| 47 |
}
|
puff/metadata.json
CHANGED
|
@@ -1,16 +1,16 @@
|
|
| 1 |
{
|
| 2 |
"character": "Jigglypuff",
|
| 3 |
"melee_enum": "JIGGLYPUFF",
|
| 4 |
-
"run_name": "puff-
|
| 5 |
-
"global_step":
|
| 6 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
"n_controller_combos": 7,
|
| 8 |
-
"model_preset": "mimic",
|
| 9 |
"pos_enc": "relpos",
|
| 10 |
-
"
|
| 11 |
-
"
|
| 12 |
-
"val_btn_f1": "91.3%",
|
| 13 |
-
"val_main_f1": "52.2%",
|
| 14 |
-
"val_shldr_f1": "91.7%",
|
| 15 |
-
"val_cdir_f1": "70.6%"
|
| 16 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"character": "Jigglypuff",
|
| 3 |
"melee_enum": "JIGGLYPUFF",
|
| 4 |
+
"run_name": "puff-20260418-relpos-xl",
|
| 5 |
+
"global_step": 29757,
|
| 6 |
+
"games_trained": 33585,
|
| 7 |
+
"val_btn_f1": "91.3%",
|
| 8 |
+
"val_main_f1": "51.7%",
|
| 9 |
+
"val_loss": "0.6766",
|
| 10 |
+
"n_params": 50590757,
|
| 11 |
"n_controller_combos": 7,
|
| 12 |
+
"model_preset": "mimic-xl",
|
| 13 |
"pos_enc": "relpos",
|
| 14 |
+
"no_self_inputs": false,
|
| 15 |
+
"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)."
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
}
|
puff/model.pt
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:65fdc96b775bf0efd240fc76cc194fd4d8f3a8ec2a5e31f5cdc429a970ad2107
|
| 3 |
+
size 556909050
|