fix: [trainers] Compatibility with the transformers v5 Trainer API

Trainer's tokenizer keyword argument was removed in transformers v5 in
favour of processing_class, so the CWE classification and description
generation trainers crashed at trainer construction. Also drop the
deprecated logging_dir training argument from all trainers.

Author: cedricbonhomme

vulntrain/trainers/classify_severity.py

@@ -204,7 +204,6 @@ def tokenize_function(elem):
         per_device_eval_batch_size=8 if "roberta" in base_model else 16,
         num_train_epochs=5,
         weight_decay=0.01,
-        logging_dir="./logs",
         logging_steps=10,
         save_total_limit=3,
         load_best_model_at_end=True,

vulntrain/trainers/classify_severity_cnvd.py

@@ -288,7 +288,6 @@ def tokenize_function(examples):
         per_device_eval_batch_size=8 if "roberta" in base_model else 16,
         num_train_epochs=5,
         weight_decay=0.01,
-        logging_dir="./logs",
         logging_steps=10,
         save_total_limit=3,
         load_best_model_at_end=True,

vulntrain/trainers/cwe_guesser_patches.py

@@ -162,7 +162,6 @@ def tokenize_function(examples):
         per_device_eval_batch_size=16,
         num_train_epochs=40,
         weight_decay=0.01,
-        logging_dir="./logs",
         logging_steps=20,
         load_best_model_at_end=True,
         push_to_hub=True,
@@ -174,7 +173,7 @@ def tokenize_function(examples):
         args=training_args,
         train_dataset=tokenized_dataset["train"],
         eval_dataset=tokenized_dataset["test"],
-        tokenizer=tokenizer,
+        processing_class=tokenizer,
         data_collator=DataCollatorWithPadding(tokenizer),
         compute_metrics=compute_metrics,
         class_weights=class_weights,

vulntrain/trainers/generation_description.py

@@ -77,7 +77,6 @@ def train(
         eval_strategy="epoch",
         save_strategy="epoch",
         load_best_model_at_end=True,
-        logging_dir="./logs",
         hub_model_id=repo_id,  # Explicitly specify HF model repo
     )
 
@@ -86,7 +85,7 @@ def train(
         args=training_args,
         train_dataset=datasets["train"],
         eval_dataset=datasets["test"],
-        tokenizer=tokenizer,
+        processing_class=tokenizer,
         data_collator=DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False),
     )