fix wrapping

Author: kjappelbaum

data/tabular/melting_points/meta.yaml

@@ -19,10 +19,10 @@ targets:
 benchmarks: []
 identifiers:
   - id: SMILES
-    type: text
+    type: SMILES
     description: SMILES
   - id: NAME
-    type: text
+    type: Other
     description: name
 license: CC BY 4.0
 links:

experiments/ablations/continued_pretrain.py

@@ -23,8 +23,8 @@ def load_model(
         dtype=dtype,
         load_in_4bit=load_in_4bit,
     )
-
-    add_new_tokens(model, tokenizer, new_tokens=add_special_tokens)
+    if add_special_tokens is not None:
+        add_new_tokens(model, tokenizer, new_tokens=add_special_tokens)
 
     target_modules = [
         "q_proj",
@@ -116,7 +116,12 @@ def formatting_prompts_func(examples):
     return dataset
 
 
-def run(data_files: List[str], train_embeddings: bool, run_name: str, batch_size: int, add_special_tokens: Optional[List[str]]=None)
+def run(data_files: List[str],  run_name: str, batch_size: int=64, add_special_tokens: Optional[List[str]]=None, train_embeddings: bool=True):
+    print(f"Data files {data_files}")
+    print(f"Run name {run_name}")
+    print(f"Batch size {batch_size}")
+    print(f"Add special tokens {add_special_tokens}")
+    print(f"Train embeddings {train_embeddings}")
     model, tokenizer = load_model(train_embeddings=train_embeddings, add_special_tokens=add_special_tokens )
 
     dataset = create_dataset(

src/chemnlp/data/sampler.py

@@ -172,6 +172,7 @@ def _wrap_identifier(self, identifier: str, value: str) -> str:
         """Wrap the identifier value with tags if wrap_identifiers is enabled."""
 
         if not self.wrap_identifiers:
+            logger.debug("Not wrapping identifiers.")
             return value
 
         identifier_type = next(
@@ -188,9 +189,11 @@ def _wrap_identifier(self, identifier: str, value: str) -> str:
         except ValueError:
             identifier_type = None
 
+        logger.debug(f'Identifier type: {identifier_type}, value: {value}')
         if identifier_type and identifier_type not in self.config.get(
             "excluded_from_wrapping", []
         ):
+            logger.debug(f"Wrapping {identifier_type} with tags.")
             return f"[BEGIN_{identifier_type}]{value}[END_{identifier_type}]"
         return value
 

src/chemnlp/data/sampler_cli.py

@@ -102,6 +102,7 @@ def process_dataset(
         "excluded_from_wrapping": ["Other"],
     }
 
+
     templates = meta["templates"]
     if benchmarking:
         templates = [t for t in templates if "<EOI>" in t]
@@ -116,7 +117,9 @@ def process_dataset(
         logger.debug(f"Processing chunk {chunk_idx} to {chunk_output_dir}")
         os.makedirs(chunk_output_dir, exist_ok=True)
 
-        sampler = TemplateSampler(df_chunk, meta, config, data_dir)
+        sampler = TemplateSampler(df_chunk, meta=meta, config=config, path_data_dir=data_dir)
+        if wrap_identifiers:
+            assert sampler.wrap_identifiers, "Wrap identifiers must be enabled in the sampler"
 
         for template_idx, template in enumerate(templates):
             print(
@@ -177,7 +180,7 @@ def main(
         benchmarking,
         additional_templates,
         use_standard_templates,
-        wrap_identifiers,
+        wrap_identifiers=wrap_identifiers,
     )