Upload ConlluTokenClassificationPipeline

config.json

@@ -27,8 +27,8 @@
   "lora_dropout": 0.05,
   "lora_r": 8,
   "lora_target_modules": [
-    "q_proj",
-    "v_proj"
+    "query",
+    "value"
   ],
   "misc_classifier_hidden_size": 512,
   "model_type": "cobald_parser",

configuration.py

@@ -26,6 +26,8 @@ class CobaldParserConfig(PretrainedConfig):
         lora_target_modules: list = None,
         **kwargs
     ):
+        print("DEBUG (encoder): use_lora:", use_lora)
+        print("DEBUG (encoder): lora_target_modules:", lora_target_modules)
         self.encoder_model_name = encoder_model_name
         self.null_classifier_hidden_size = null_classifier_hidden_size
         self.consecutive_null_limit = consecutive_null_limit

encoder.py

@@ -47,8 +47,15 @@ class WordTransformerEncoder(nn.Module):
                 bias="none",
                 task_type="SEQ_CLS"
             )
+            print("DEBUG: model class =", type(self.model))
+            for name, module in self.model.named_modules():
+                if "proj" in name:
+                    print("DEBUG: found module", name, "->", module)
             self.model = get_peft_model(self.model, lora_config)
-            print(f"LoRA enabled: r={lora_r}, alpha={lora_alpha}, target_modules={lora_target_modules}")
+            print("LoRA enabled! Model type:", type(self.model))
+            for name, param in self.model.named_parameters():
+                if "lora" in name:
+                    print("LoRA param:", name, param.shape)
 
     def forward(self, words: list[list[str]]) -> Tensor:
         """

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a48711fc2496833a9e5a00ca3563b0ed3eab04d3e4c0c3cc3cc49754134ae349
+oid sha256:37137a12604aa1ee98f7cc4627a3ed76c63cdb9eb9d2fd02c4a73aaf17325dae
 size 1134190536

modeling_parser.py

@@ -24,7 +24,12 @@ class CobaldParser(PreTrainedModel):
         super().__init__(config)
 
         self.encoder = WordTransformerEncoder(
-            model_name=config.encoder_model_name
+            model_name=config.encoder_model_name, 
+            use_lora=getattr(config, "use_lora", False),
+            lora_r=getattr(config, "lora_r", 8),
+            lora_alpha=getattr(config, "lora_alpha", 16),
+            lora_dropout=getattr(config, "lora_dropout", 0.05),
+            lora_target_modules=getattr(config, "lora_target_modules", None),
         )
         embedding_size = self.encoder.get_embedding_size()