Improve the 3-class sentence-based model.

In the previous training we inadvertently used a sequence length of just `20` tokens,
which is what `AutoConfig.from_pretrained defaults to. This model has been trained
with up to 512 tokens, better utilizing Roberta's capabilities.

F1 macro score is now `0.6951`.

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "./artifacts/model-36sxpfct-2744-0.5888:v0",
+  "_name_or_path": "./artifacts/model-1qe3kbqj-2742-0.6951:v0",
   "architectures": [
     "RobertaForSequenceClassification"
   ],
@@ -24,6 +24,7 @@
     "intermediate": 1
   },
   "layer_norm_eps": 1e-05,
+  "max_length": 512,
   "max_position_embeddings": 514,
   "model_type": "roberta",
   "num_attention_heads": 12,

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:56ad453cd87d0c3835c6811c196e8dfbb0f3063cfa59226d8dbc686bdef31f16
+oid sha256:26854ee4fa633135cd6bc6f190398ae0cca099a8d3e9adeb54c94bb61a536530
 size 498675309