[transformers] Prompt masking

src/sparseml/transformers/finetune/data/base.py

@@ -41,6 +41,7 @@ class TextGenerationDataset(RegistryMixin):
     """
 
     PROMPT_KEY = "prompt"
+    MASK_KEY = "mask"
 
     def __init__(
         self,
@@ -125,6 +126,7 @@ def tokenize_fn(data):
                 padding=self.padding,
                 max_length=self.max_seq_length,
                 truncation=True,
+                return_offsets_mapping=True,
             )
 
             # store unpadded prompt so we can mask out correct number of elements
@@ -156,16 +158,29 @@ def group_text_fn(data):
         def label_fn(data):
             # if the dataset uses prompts, mask them out so they don't contribute
             # to the loss calculation
+            labels = data["input_ids"].copy()
+            if "offset_mapping" in data:
+                offset_mapping = data["offset_mapping"]
+                # get the character level mask
+                mask = data.get("mask")
+                if mask is not None:
+                    for i, (start, end) in enumerate(offset_mapping):
+                        # if any char is to be filtered
+                        if "0" in mask[start:end]:
+                            labels[i] = LABELS_MASK_VALUE
+
             prompt_len = 0
             if self.PROMPT_KEY in data:
                 prompt_len = len(data[self.PROMPT_KEY])
-            data["labels"] = data["input_ids"].copy()
+
+            data["labels"] = labels
             data["labels"][:prompt_len] = [LABELS_MASK_VALUE] * prompt_len
 
             # mask out padding in the labels as well
             padding = len(data["attention_mask"]) - sum(data["attention_mask"])
             if padding > 0:
                 data["labels"][-padding:] = [LABELS_MASK_VALUE] * padding
+
             return data
 
         if raw_dataset is None:
@@ -209,8 +224,6 @@ def label_fn(data):
             load_from_cache_file=not self.data_args.overwrite_cache,
             desc="Adding labels",
         )
-        print(dataset.column_names)
-
         return dataset
 
     def map(
@@ -229,5 +242,4 @@ def map(
             kwargs.pop("num_proc", None)
             kwargs.pop("load_from_cache_file", None)
             kwargs.pop("desc", None)
-
         return dataset.map(**kwargs)

src/sparseml/transformers/finetune/data/custom.py

@@ -91,7 +91,6 @@ def get_raw_dataset(self, *_ignore, **__ignore) -> Union[DatasetDict, Dataset]:
                 num_proc=self.data_args.preprocessing_num_workers,
                 desc="Removing unneeded columns",
             )
-
         return raw_dataset
 
     def get_remove_columns_from_dataset(
@@ -108,5 +107,7 @@ def get_remove_columns_from_dataset(
             remove_columns.remove(self.text_column)
         if self.PROMPT_KEY in remove_columns:
             remove_columns.remove(self.PROMPT_KEY)
+        if self.MASK_KEY in remove_columns:
+            remove_columns.remove(self.MASK_KEY)
 
         return list(remove_columns)

src/sparseml/transformers/utils/helpers.py

@@ -54,6 +54,7 @@
     "ALL_TASK_NAMES",
     "create_fake_dataloader",
     "POSSIBLE_TOKENIZER_FILES",
+    "generate_mask",
     "download_repo_from_huggingface_hub",
     "download_model_directory",
 ]
@@ -544,6 +545,56 @@ def fetch_recipe_path(target: str):
     return recipe_path
 
 
+def generate_mask(string: str, response: str, prompt: str = "") -> str:
+    """
+    Generate a mask based on provided prompt and response strings to obscure
+    characters in the input string. Prompt will be masked and string in response
+    will be kept represented by 0 - remove and 1 - keep.
+    By default, non-reponse wrapped strings will be matched with 0
+
+    Args:
+    :param string: The input string to be masked.
+    :param prompt: The prompt string to identify characters to obscure.
+    :param response: The response string to identify characters to keep visible.
+
+    Returns:
+        str: A string representing the mask where '1' indicates visible
+        characters and '0' indicates obscured characters.
+
+    """
+
+    mask = ["1"] * len(string)
+    is_prompt = False if string.startswith(response) else True
+    counter = 0
+    for i, char in enumerate(string):
+        if is_prompt:
+            mask[i] = "0"
+
+        if counter > 0:
+            if not is_prompt and len(prompt) > 1 and char == prompt[counter]:
+                counter += 1
+            elif is_prompt and char == response[counter]:
+                counter += 1
+            else:
+                counter = 0
+
+        if len(prompt) > 0 and counter == len(prompt) and not is_prompt:
+            mask[i - counter + 1 : i + 1] = ["0"] * counter
+
+            counter = 0
+            is_prompt = True
+
+        if counter == len(response) and is_prompt:
+            mask[i - counter + 1 : i + 1] = ["1"] * counter
+
+            counter = 0
+            is_prompt = False
+
+        if prompt.startswith(char) or response.startswith(char):
+            counter = 1
+    return "".join(mask)
+
+
 def download_repo_from_huggingface_hub(repo_id, **kwargs):
     """
     Download relevant model files from the Hugging Face Hub

src/sparseml/transformers/utils/preprocessing_functions.py

@@ -14,6 +14,7 @@
 
 from typing import Dict
 
+from sparseml.transformers.utils.helpers import generate_mask
 from sparsezoo.utils.registry import RegistryMixin
 
 
@@ -26,4 +27,7 @@ def custom_evolved_codealpaca_dataset(data: Dict):
     PROMPT_DICT = """[Instruction]:\n{instruction}\n\n[Response]:"""
     data["prompt"] = PROMPT_DICT.format_map(data)
     data["text"] = data["prompt"] + data["output"]
+    data["mask"] = generate_mask(
+        data["text"], prompt="[Instruction]", censor="[Response]"
+    )
     return data

tests/sparseml/transformers/finetune/test_finetune.py

@@ -33,6 +33,7 @@
     oneshot,
     train,
 )
+from sparseml.transformers.utils.helpers import generate_mask
 
 
 def test_oneshot_and_finetune(tmp_path: Path):
@@ -322,3 +323,35 @@ def test_oneshot_with_modifier_object(tmp_path: Path):
         splits=splits,
         oneshot_device=device,
     )
+
+
+def test_finetune_wout_recipe_with_mask(tmp_path: Path):
+    recipe_str = None
+    model = "Xenova/llama2.c-stories15M"
+    device = "cuda:0"
+    if not torch.cuda.is_available():
+        device = "cpu"
+    dataset = "open_platypus"
+    concatenate_data = False
+    output_dir = tmp_path
+    max_steps = 50
+    splits = "train"
+
+    def preprocessing_func(example):
+        example["text"] = "[foo]" + example["text"] + "[bar] mask this"
+        example["mask"] = generate_mask(
+            example["text"], response="[bar]", prompt="[foo]"
+        )
+        return example
+
+    train(
+        model=model,
+        dataset=dataset,
+        output_dir=output_dir,
+        recipe=recipe_str,
+        max_steps=max_steps,
+        concatenate_data=concatenate_data,
+        splits=splits,
+        oneshot_device=device,
+        preprocessing_func=preprocessing_func,
+    )

tests/sparseml/transformers/utils/test_helpers.py

@@ -21,6 +21,7 @@
 from accelerate import init_empty_weights
 from sparseml.transformers.utils.helpers import (
     create_fake_dataloader,
+    generate_mask,
     infer_recipe_from_model_path,
     is_transformer_model,
     resolve_recipe_file,
@@ -166,3 +167,55 @@ def test_save_zoo_directory(tmp_path, stub):
     assert zoo_model.validate(minimal_validation=True, validate_onnxruntime=False)
     shutil.rmtree(path_to_training_outputs)
     shutil.rmtree(save_dir)
+
+
+@pytest.mark.parametrize(
+    "string, response, prompt, expected_mask",
+    [
+        (
+            ("[foo]hello\n\n" "[bar]world"),
+            "[bar]",
+            "[foo]",
+            ("000000000000" "1111111111"),
+        ),
+        (
+            (
+                "[Instruction]python is\n\n"  # 24
+                "[Response]great\n\n"  # 17
+                "[Instruction]What about Java"  # 28
+                "[Response]Meh"  # 13
+            ),
+            "[Response]",
+            "[Instruction]",
+            (
+                "000000000000000000000000"  # 24
+                "11111111111111111"  # 17
+                "0000000000000000000000000000"  # 28
+                "1111111111111"  # 13
+            ),
+        ),
+        (
+            ("[foo]hello\n\n" "[bar]world"),
+            "[bar]",
+            None,
+            ("000000000000" "1111111111"),
+        ),
+        (
+            ("hello\n\n" "[bar]world"),
+            "[bar]",
+            None,
+            ("0000000" "1111111111"),
+        ),
+        (
+            ("[bar]world" "[foo]hello\n\n" "[bar]world"),
+            "[bar]",
+            "[foo]",
+            ("1111111111" "000000000000" "1111111111"),
+        ),
+    ],
+)
+def test_generate_mask(string, response, prompt, expected_mask):
+    if prompt is not None:
+        assert generate_mask(string, response, prompt) == expected_mask
+    else:
+        assert generate_mask(string, response) == expected_mask