Replace epoch-4 word-level checkpoint with epoch-7 BPE-tokenizer run

Old checkpoint + config preserved at tag epoch4-wordlevel. New: model.mpk (epoch 7, lowest valid loss 3.999), config.json, tokenizer.json (byte-level BPE, 32k vocab).

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 checkpoint/model-4.mpk filter=lfs diff=lfs merge=lfs -text
 checkpoint/optim-4.mpk filter=lfs diff=lfs merge=lfs -text
+model.mpk filter=lfs diff=lfs merge=lfs -text

checkpoint/optim-4.mpk

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:98ad42f2bbaed8bb8d09c31c546774759368c33b3c54de27b6f404b7bf1f071b
-size 540030423

checkpoint/scheduler-4.mpk

@@ -1 +0,0 @@
-��metadata��float�half::binary16::f16�int�i16�format�aburn_core::record::file::NamedMpkFileRecorder<burn_core::record::settings::HalfPrecisionSettings>�version�0.21.0�settings�HalfPrecisionSettings�item��_b�

config.json

@@ -2,7 +2,7 @@
   "model": {
     "embedding_config": {
       "context_size": 1024,
-      "vocab_size": 32002,
+      "vocab_size": 32000,
       "d_model": 768,
       "dropout_prob": 0.1
     },
@@ -27,9 +27,11 @@
     "cautious_weight_decay": false,
     "amsgrad": false
   },
+  "tokenizer_kind": "bpe",
+  "tokenizer_path": "tokenizer.json",
   "num_epochs": 200,
-  "batch_size": 20,
-  "num_workers": 32,
+  "batch_size": 6,
+  "num_workers": 6,
   "seed": 123456,
   "learning_rate": 0.0001
 }

checkpoint/model-4.mpk → model.mpk

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b228b2fede7890d0161dcae6ce40f3c0b2fe40c83ce5ae87abf85ba6d063be0b
-size 270016715
+oid sha256:b72b4708e03ae67ff9ac0fc0a1e779f10a25a03ff70dfb9526893393fe027654
+size 270010567