Add desc to map/filter (#1162)

* Add desc to map/filter

* update descriptions

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

src/axolotl/cli/__init__.py

@@ -410,7 +410,10 @@ def ultra_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
     for i, data_set in enumerate(train_datasets):
         _type = cfg.datasets[i]["type"]
         ds_type_fn = locals()[_type]
-        train_datasets[i] = data_set.map(ds_type_fn)
+        train_datasets[i] = data_set.map(
+            ds_type_fn,
+            desc="Mapping RL Dataset",
+        )
     train_dataset = concatenate_datasets(train_datasets)
 
     # eval_dataset = eval_dataset.map(intel_apply_chatml)

src/axolotl/datasets.py

@@ -57,6 +57,7 @@ def process(self, dataset):
             num_proc=num_proc,
             remove_columns=features,
             keep_in_memory=self.keep_in_memory,
+            desc="Tokenizing Prompts",
             **map_kwargs,
         )
 

src/axolotl/utils/data.py

@@ -792,6 +792,7 @@ def load_pretraining_dataset(path, tokenizer, cfg, name=None, max_tokens=2048, s
         # remove all the existing columns after mapping since they end up having
         # a different length than the encoded/tokenized column
         remove_columns=dataset.features.keys(),
+        desc="Encoding Pretraining",
     )
     return dataset
 

src/axolotl/utils/trainer.py

@@ -134,33 +134,38 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
             drop_long,
             num_proc=cfg.dataset_processes,
             load_from_cache_file=not cfg.is_preprocess,
+            desc="Dropping Long Sequences",
         )
         if eval_dataset:
             eval_dataset = eval_dataset.filter(
                 drop_long,
                 num_proc=cfg.dataset_processes,
                 load_from_cache_file=not cfg.is_preprocess,
+                desc="Dropping Long Sequences",
             )
 
         if cfg.group_by_length:
             train_dataset = train_dataset.map(
                 add_length,
                 num_proc=cfg.dataset_processes,
                 load_from_cache_file=not cfg.is_preprocess,
+                desc="Group By Length",
             )
 
         if cfg.sample_packing:
             train_dataset = train_dataset.map(
                 add_position_ids,
                 num_proc=cfg.dataset_processes,
                 load_from_cache_file=not cfg.is_preprocess,
+                desc="Add position_id column (Sample Packing)",
             )
             if cfg.eval_sample_packing is not False:
                 if eval_dataset:
                     eval_dataset = eval_dataset.map(
                         add_position_ids,
                         num_proc=cfg.dataset_processes,
                         load_from_cache_file=not cfg.is_preprocess,
+                        desc="Add position_id column (Sample Packing)",
                     )
 
     return train_dataset, eval_dataset
@@ -169,9 +174,13 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
 def process_pretraining_datasets_for_packing(train_dataset, sequence_len):
     drop_long = partial(drop_long_seq, sequence_len=sequence_len)
 
-    train_dataset = train_dataset.filter(drop_long)
+    train_dataset = train_dataset.filter(
+        drop_long,
+        desc="Dropping Long Sequences",
+    )
     train_dataset = train_dataset.map(
         add_position_ids,
+        desc="Add position_id column (Pretraining Sample Packing)",
     )
     return train_dataset