cehrgpt 0.0.2__tar.gz → 0.1.0__tar.gz

{cehrgpt-0.0.2 → cehrgpt-0.1.0}/.gitignore

@@ -4,6 +4,9 @@
 venv*
 dist/*
 
+*png
+*json
+
 *ipynb_checkpoints/
 *h5
 *logs

{cehrgpt-0.0.2/src/cehrgpt.egg-info → cehrgpt-0.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: cehrgpt
-Version: 0.0.2
+Version: 0.1.0
 Summary: CEHR-GPT: Generating Electronic Health Records with Chronological Patient Timelines
 Author-email: Chao Pang <chaopang229@gmail.com>, Xinzhuo Jiang <xj2193@cumc.columbia.edu>, Krishna Kalluri <kk3326@cumc.columbia.edu>, Elise Minto <em3697@cumc.columbia.edu>, Jason Patterson <jp3477@cumc.columbia.edu>, Nishanth Parameshwar Pavinkurve <np2689@cumc.columbia.edu>, Karthik Natarajan <kn2174@cumc.columbia.edu>
 License: MIT License
@@ -12,13 +12,14 @@ Classifier: Programming Language :: Python :: 3
 Requires-Python: >=3.10.0
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: cehrbert==1.3.8
+Requires-Dist: cehrbert==1.4.1
+Requires-Dist: cehrbert_data==0.0.7
 Requires-Dist: openai==1.54.3
 Requires-Dist: optuna==4.0.0
-Requires-Dist: transformers==4.40.0
+Requires-Dist: transformers==4.44.0
 Requires-Dist: tokenizers==0.19.0
 Requires-Dist: peft==0.10.0
-Requires-Dist: trl==0.11.4
+Requires-Dist: lightgbm
 Provides-Extra: dev
 Requires-Dist: pre-commit; extra == "dev"
 Requires-Dist: pytest; extra == "dev"
@@ -29,6 +30,7 @@ Requires-Dist: hypothesis; extra == "dev"
 Requires-Dist: black; extra == "dev"
 Provides-Extra: flash-attn
 Requires-Dist: flash_attn; extra == "flash-attn"
+Dynamic: license-file
 
 # CEHRGPT
 

{cehrgpt-0.0.2 → cehrgpt-0.1.0}/pyproject.toml

@@ -28,13 +28,14 @@ classifiers = [
 ]
 
 dependencies = [
-    "cehrbert==1.3.8",
+    "cehrbert==1.4.1",
+    "cehrbert_data==0.0.7",
     "openai==1.54.3",
     "optuna==4.0.0",
-    "transformers==4.40.0",
+    "transformers==4.44.0",
     "tokenizers==0.19.0",
     "peft==0.10.0",
-    "trl==0.11.4",
+    "lightgbm",
 ]
 
 [tool.setuptools_scm]

{cehrgpt-0.0.2 → cehrgpt-0.1.0}/scripts/level_three_evaluation.sh

@@ -29,7 +29,8 @@ python -u -m cehrbert_data.prediction_cohorts.cad_cabg_cohort \
     -dl 1985-01-01 -du 2023-12-31 \
     -l 18 -u 100 -ow 360 -ps 0 -pw 360 -f \
     --att_type cehr_bert \
-    --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv
+    --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv \
+    --is_remove_index_prediction_starts
 
 # Run Predictions on CAD CABG
 echo "Run predictions on cad_cabg"
@@ -56,9 +57,10 @@ python -u -m cehrbert_data.prediction_cohorts.hf_readmission \
   -c hf_readmission_bow \
   -i "$OMOP_FOLDER" \
   -o "$OMOP_FOLDER/cohorts/hf_readmission" \
-  -dl 1985-01-01 -du 2023-12-31 -l 18 -u 100 -ow 360 -ps 0 -pw 30 -f \
+  -dl 1985-01-01 -du 2023-12-31 -l 18 -u 100 -ow 360 -ps 1 -pw 30 -f \
   --att_type cehr_bert \
-  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv
+  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv \
+  --is_remove_index_prediction_starts
 
 # Run predictions on HF Readmission
 echo "Run predictions on hf_readmission"
@@ -85,9 +87,10 @@ python -u -m cehrbert_data.prediction_cohorts.copd_readmission \
   -c copd_readmission_bow \
   -i "$OMOP_FOLDER" \
   -o "$OMOP_FOLDER/cohorts/copd_readmission" \
-  -dl 1985-01-01 -du 2023-12-31 -l 18 -u 100 -ow 720 -ps 0 -pw 360 -f \
+  -dl 1985-01-01 -du 2023-12-31 -l 18 -u 100 -ow 360 -ps 1 -pw 30 -f \
   --att_type cehr_bert \
-  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv
+  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv \
+  --is_remove_index_prediction_starts
 
 # Run predictions on COPD Readmission
 echo "Run predictions on copd_readmission"
@@ -145,7 +148,8 @@ python -u -m cehrbert_data.prediction_cohorts.afib_ischemic_stroke \
   -o "$OMOP_FOLDER/cohorts/afib_ischemic_stroke" \
   -dl 1985-01-01 -du 2023-12-31 -l 18 -u 100 -ow 720 -ps 0 -pw 360 -f \
   --att_type cehr_bert \
-  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv
+  --ehr_table_list condition_occurrence procedure_occurrence drug_exposure -iv \
+  --is_remove_index_prediction_starts
 
 # Run predictions on AFIB Ischemic Stroke
 echo "Run predictions on afib_ischemic_stroke"

{cehrgpt-0.0.2 → cehrgpt-0.1.0}/src/cehrgpt/data/hf_cehrgpt_dataset.py

@@ -1,9 +1,10 @@
-from typing import Union
+from typing import Optional, Union
 
 from cehrbert.data_generators.hf_data_generator.hf_dataset import (
     FINETUNING_COLUMNS,
     apply_cehrbert_dataset_mapping,
 )
+from cehrbert.data_generators.hf_data_generator.meds_utils import CacheFileCollector
 from cehrbert.runners.hf_runner_argument_dataclass import DataTrainingArguments
 from datasets import Dataset, DatasetDict
 
@@ -31,16 +32,25 @@ def create_cehrgpt_pretraining_dataset(
     dataset: Union[Dataset, DatasetDict],
     cehrgpt_tokenizer: CehrGptTokenizer,
     data_args: DataTrainingArguments,
-) -> Dataset:
+    cache_file_collector: Optional[CacheFileCollector] = None,
+) -> Union[Dataset, DatasetDict]:
     required_columns = TRANSFORMER_COLUMNS + CEHRGPT_COLUMNS
+    # TODO: temp solution, this column is mixed typed and causes an issue when transforming the data
+    if not data_args.streaming:
+        if isinstance(dataset, DatasetDict):
+            all_columns = dataset["train"].column_names
+        else:
+            all_columns = dataset.column_names
+        if "visit_concept_ids" in all_columns:
+            dataset.remove_columns(["visit_concept_ids"])
     dataset = apply_cehrbert_dataset_mapping(
         dataset,
         HFCehrGptTokenizationMapping(cehrgpt_tokenizer),
         num_proc=data_args.preprocessing_num_workers,
         batch_size=data_args.preprocessing_batch_size,
         streaming=data_args.streaming,
+        cache_file_collector=cache_file_collector,
     )
-
     if not data_args.streaming:
         if isinstance(dataset, DatasetDict):
             all_columns = dataset["train"].column_names
@@ -56,8 +66,17 @@ def create_cehrgpt_finetuning_dataset(
     dataset: Union[Dataset, DatasetDict],
     cehrgpt_tokenizer: CehrGptTokenizer,
     data_args: DataTrainingArguments,
-) -> Dataset:
+    cache_file_collector: Optional[CacheFileCollector] = None,
+) -> Union[Dataset, DatasetDict]:
     required_columns = TRANSFORMER_COLUMNS + CEHRGPT_COLUMNS + FINETUNING_COLUMNS
+    # TODO: temp solution, this column is mixed typed and causes an issue when transforming the data
+    if not data_args.streaming:
+        if isinstance(dataset, DatasetDict):
+            all_columns = dataset["train"].column_names
+        else:
+            all_columns = dataset.column_names
+        if "visit_concept_ids" in all_columns:
+            dataset.remove_columns(["visit_concept_ids"])
     mapping_functions = [
         HFFineTuningMapping(cehrgpt_tokenizer),
     ]
@@ -68,6 +87,7 @@ def create_cehrgpt_finetuning_dataset(
             num_proc=data_args.preprocessing_num_workers,
             batch_size=data_args.preprocessing_batch_size,
             streaming=data_args.streaming,
+            cache_file_collector=cache_file_collector,
         )
 
     if not data_args.streaming:

{cehrgpt-0.0.2 → cehrgpt-0.1.0}/src/cehrgpt/data/hf_cehrgpt_dataset_collator.py

@@ -1,5 +1,5 @@
 import random
-from typing import Any, Dict
+from typing import Any, Dict, Optional
 
 import numpy as np
 import torch
@@ -105,9 +105,12 @@ class CehrGptDataCollator:
             self._try_reverse_tensor(self._convert_to_tensor(example["input_ids"]))
             for example in examples
         ]
+
         batch_attention_mask = [
             self._try_reverse_tensor(
-                torch.ones_like(
+                self._convert_to_tensor(example["attention_mask"]).to(torch.float)
+                if "attention_mask" in example
+                else torch.ones_like(
                     self._convert_to_tensor(example["input_ids"]), dtype=torch.float
                 )
             )
@@ -128,16 +131,40 @@ class CehrGptDataCollator:
         )
         assert batch["input_ids"].shape[1] <= self.max_length
         assert batch["attention_mask"].shape[1] <= self.max_length
+        assert batch["attention_mask"].shape[1] == batch["input_ids"].shape[1], (
+            f'batch["attention_mask"].shape[1]: {batch["attention_mask"].shape[1]}, '
+            f'batch["input_ids"].shape[1]: {batch["input_ids"].shape[1]}'
+        )
+        assert batch["input_ids"].max() < self.tokenizer.vocab_size, (
+            f"batch['input_ids'].max(): {batch['input_ids'].max()} must be smaller than "
+            f"self.tokenizer.vocab_size: {self.tokenizer.vocab_size}. "
+            f"batch['input_ids']: {batch['input_ids']} "
+        )
 
-        if self.pretraining:
-            batch["labels"] = self._try_reverse_tensor(
+        if "position_ids" in examples[0]:
+            batch_position_ids = [
+                self._try_reverse_tensor(
+                    self._convert_to_tensor(example["position_ids"])
+                )
+                for example in examples
+            ]
+            # Pad sequences to the max length in the batch
+            batch["position_ids"] = self._try_reverse_tensor(
                 pad_sequence(
-                    batch_input_ids,
+                    batch_position_ids,
                     batch_first=True,
-                    padding_value=-100,
+                    padding_value=self.max_length,
                 ).to(torch.int64)
             )
 
+        if self.pretraining:
+            batch["labels"] = torch.where(
+                (batch["input_ids"] != self.tokenizer.pad_token_id)
+                & batch["attention_mask"].to(torch.bool),
+                batch["input_ids"],
+                -100,
+            )
+
         if self.use_sub_time_tokenization:
             time_token_indicators = torch.isin(batch["input_ids"], self.time_tokens)
             masked_tokens = batch["input_ids"].clone()
@@ -170,7 +197,7 @@ class CehrGptDataCollator:
         if self.include_values:
             batch_value_indicators = [
                 self._try_reverse_tensor(
-                    self._convert_to_tensor(example["value_indicators"])
+                    self._convert_to_tensor(example["value_indicators"]).to(torch.bool)
                 )
                 for example in examples
             ]
@@ -178,7 +205,6 @@ class CehrGptDataCollator:
                 self._try_reverse_tensor(self._convert_to_tensor(example["values"]))
                 for example in examples
             ]
-
             batch["value_indicators"] = self._try_reverse_tensor(
                 pad_sequence(
                     batch_value_indicators, batch_first=True, padding_value=False
@@ -200,41 +226,58 @@ class CehrGptDataCollator:
                     batch["value_indicators"], batch["values"].clone(), -100
                 )
 
+        bz = len(examples)
         if "person_id" in examples[0]:
-            batch["person_id"] = torch.cat(
-                [
-                    self._convert_to_tensor(example["person_id"]).reshape(-1, 1)
-                    for example in examples
-                ],
-                dim=0,
-            ).to(torch.int32)
+            batch["person_id"] = (
+                torch.cat(
+                    [
+                        self._convert_to_tensor(example["person_id"]).reshape(-1, 1)
+                        for example in examples
+                    ],
+                    dim=0,
+                )
+                .to(torch.int32)
+                .reshape(bz, -1)
+            )
 
         if "index_date" in examples[0]:
             batch["index_date"] = torch.cat(
                 [
-                    self._convert_to_tensor(example["index_date"]).reshape(-1, 1)
+                    torch.tensor(example["index_date"], dtype=torch.float64).reshape(
+                        -1, 1
+                    )
                     for example in examples
                 ],
                 dim=0,
-            ).to(torch.float32)
+            ).reshape(bz, -1)
 
         if "age_at_index" in examples[0]:
-            batch["age_at_index"] = torch.cat(
-                [
-                    self._convert_to_tensor(example["age_at_index"]).reshape(-1, 1)
-                    for example in examples
-                ],
-                dim=0,
-            ).to(torch.float32)
+            batch["age_at_index"] = (
+                torch.cat(
+                    [
+                        self._convert_to_tensor(example["age_at_index"]).reshape(-1, 1)
+                        for example in examples
+                    ],
+                    dim=0,
+                )
+                .to(torch.float32)
+                .reshape(bz, -1)
+            )
 
         if "classifier_label" in examples[0]:
-            batch["classifier_label"] = torch.cat(
-                [
-                    self._convert_to_tensor(example["classifier_label"]).reshape(-1, 1)
-                    for example in examples
-                ],
-                dim=0,
-            ).to(torch.float32)
+            batch["classifier_label"] = (
+                torch.cat(
+                    [
+                        self._convert_to_tensor(example["classifier_label"]).reshape(
+                            -1, 1
+                        )
+                        for example in examples
+                    ],
+                    dim=0,
+                )
+                .to(torch.float32)
+                .reshape(bz, -1)
+            )
 
         return batch
 
@@ -273,53 +316,69 @@ class CehrGptDataCollator:
             record["input_ids"] = self._convert_to_tensor(sorted_input_ids)
         return record
 
-    def generate_start_end_index(self, record: Dict[str, Any]) -> Dict[str, Any]:
+    def generate_start_end_index(
+        self, record: Dict[str, Any], max_length_allowed: Optional[int] = None
+    ) -> Dict[str, Any]:
         """Adding the start and end indices to extract a portion of the patient sequence."""
         # concept_ids will be used to for time to event predictions and identifying the visit starts
+        max_length_allowed = (
+            self.max_length if max_length_allowed is None else max_length_allowed
+        )
+        sample_packing = getattr(self, "sample_packing", False)
         input_ids = record["input_ids"]
         if isinstance(input_ids, torch.Tensor):
             input_ids = input_ids.detach().tolist()
         concept_ids = self.tokenizer.decode(input_ids, skip_special_tokens=False)
         seq_length = len(record["input_ids"])
-        new_max_length = self.max_length - 1  # Subtract one for the [END] token
+
+        # Subtract one for the [END] token when sample_packing is not enabled
+        new_max_length = (
+            max_length_allowed if sample_packing else max_length_allowed - 1
+        )
+
+        if self.include_ttv_prediction:
+            record["time_to_visits"] = torch.concat(
+                [self._convert_to_tensor(self._convert_time_to_event(concept_ids))]
+            )
 
         # Return the record directly if the actual sequence length is less than the max sequence
         if seq_length <= new_max_length:
-            record["input_ids"] = torch.concat(
-                [
-                    self._convert_to_tensor(record["input_ids"]),
-                    self._convert_to_tensor([self.tokenizer.end_token_id]),
-                ]
-            )
-            if self.include_values:
-                record["value_indicators"] = torch.concat(
-                    [
-                        self._convert_to_tensor(record["value_indicators"]),
-                        self._convert_to_tensor([False]),
-                    ]
-                ).to(torch.bool)
-                record["values"] = torch.concat(
-                    [
-                        self._convert_to_tensor(record["values"]),
-                        self._convert_to_tensor([self.tokenizer.pad_value_token_id]),
-                    ]
-                )
-            if self.include_ttv_prediction:
-                record["time_to_visits"] = torch.concat(
+            if not sample_packing:
+                record["input_ids"] = torch.concat(
                     [
-                        self._convert_to_tensor(
-                            self._convert_time_to_event(concept_ids)
-                        ),
-                        self._convert_to_tensor([-100.0]),
+                        self._convert_to_tensor(record["input_ids"]),
+                        self._convert_to_tensor([self.tokenizer.end_token_id]),
                     ]
                 )
+                if self.include_values:
+                    record["value_indicators"] = torch.concat(
+                        [
+                            self._convert_to_tensor(record["value_indicators"]),
+                            self._convert_to_tensor([False]),
+                        ]
+                    ).to(torch.bool)
+                    record["values"] = torch.concat(
+                        [
+                            self._convert_to_tensor(record["values"]),
+                            self._convert_to_tensor(
+                                [self.tokenizer.pad_value_token_id]
+                            ),
+                        ]
+                    )
+                if self.include_ttv_prediction:
+                    record["time_to_visits"] = torch.concat(
+                        [
+                            record["time_to_visits"],
+                            self._convert_to_tensor([-100.0]),
+                        ]
+                    )
 
             return record
 
         if self.pretraining:
             # There is a 50% chance we randomly slice out a portion of the patient history and update the demographic
             # prompt depending on the new starting point
-            if random.random() < 0.5:
+            if random.random() < 0.5 and not sample_packing:
                 start_index, end_index, demographic_tokens = random_slice_gpt_sequence(
                     concept_ids, new_max_length
                 )
@@ -351,6 +410,11 @@ class CehrGptDataCollator:
                     break
 
             record["input_ids"] = record["input_ids"][0:end_index]
+
+            # We want to make sure we take the subset of attention_mask in sample packing if this field is available
+            if sample_packing and "attention_mask" in record:
+                record["attention_mask"] = record["attention_mask"][0:end_index]
+
             if self.include_values:
                 record["value_indicators"] = self._convert_to_tensor(
                     record["value_indicators"][0:end_index]
@@ -364,7 +428,7 @@ class CehrGptDataCollator:
                 )
             return record
         else:
-            if self.include_demographics:
+            if self.include_demographics and not sample_packing:
                 # We employ a left truncation strategy, where the most recent patient history is reserved for fine-tuning
                 demographic_prompts_at_visits = collect_demographic_prompts_at_visits(
                     concept_ids
@@ -427,6 +491,10 @@ class CehrGptDataCollator:
                     current_token = record["input_ids"][i]
                     if current_token == self.vs_token_id:
                         record["input_ids"] = record["input_ids"][i:end_index]
+                        if sample_packing and "attention_mask" in record:
+                            record["attention_mask"] = record["attention_mask"][
+                                i:end_index
+                            ]
                         if self.include_values:
                             record["value_indicators"] = record["value_indicators"][
                                 i:end_index
@@ -442,6 +510,10 @@ class CehrGptDataCollator:
             # We simply take the last new_max_length number of tokens from the patient sequence
             if len(record["input_ids"]) > new_max_length:
                 record["input_ids"] = record["input_ids"][-new_max_length:]
+                if sample_packing and "attention_mask" in record:
+                    record["attention_mask"] = record["attention_mask"][
+                        -new_max_length:
+                    ]
                 if self.include_values:
                     record["value_indicators"] = record["value_indicators"][
                         -new_max_length:
@@ -452,31 +524,135 @@ class CehrGptDataCollator:
                         -new_max_length:
                     ]
 
-            # Finally we add the end token to the end of the sequence
-            record["input_ids"] = torch.concat(
-                [
-                    self._convert_to_tensor(record["input_ids"]),
-                    self._convert_to_tensor([self.tokenizer.end_token_id]),
-                ]
-            )
-            if self.include_values:
-                record["value_indicators"] = torch.concat(
-                    [
-                        self._convert_to_tensor(record["value_indicators"]),
-                        self._convert_to_tensor([False]),
-                    ]
-                ).to(torch.bool)
-                record["values"] = torch.concat(
-                    [
-                        self._convert_to_tensor(record["values"]),
-                        self._convert_to_tensor([self.tokenizer.pad_value_token_id]),
-                    ]
-                )
-            if self.include_ttv_prediction:
-                record["time_to_visits"] = torch.concat(
+            if not sample_packing:
+                # Finally we add the end token to the end of the sequence
+                record["input_ids"] = torch.concat(
                     [
-                        record["time_to_visits"],
-                        self._convert_to_tensor([-100.0]),
+                        self._convert_to_tensor(record["input_ids"]),
+                        self._convert_to_tensor([self.tokenizer.end_token_id]),
                     ]
                 )
+                if self.include_values:
+                    record["value_indicators"] = torch.concat(
+                        [
+                            self._convert_to_tensor(record["value_indicators"]),
+                            self._convert_to_tensor([False]),
+                        ]
+                    ).to(torch.bool)
+                    record["values"] = torch.concat(
+                        [
+                            self._convert_to_tensor(record["values"]),
+                            self._convert_to_tensor(
+                                [self.tokenizer.pad_value_token_id]
+                            ),
+                        ]
+                    )
+                if self.include_ttv_prediction:
+                    record["time_to_visits"] = torch.concat(
+                        [
+                            record["time_to_visits"],
+                            self._convert_to_tensor([-100.0]),
+                        ]
+                    )
             return record
+
+
+class SamplePackingCehrGptDataCollator(CehrGptDataCollator):
+    def __init__(self, max_tokens, max_position_embeddings, *args, **kwargs):
+        self.max_tokens_per_batch = max_tokens
+        self.max_position_embeddings = max_position_embeddings
+        self.sample_packing = True
+        self.add_end_token_in_sample_packing = kwargs.pop(
+            "add_end_token_in_sample_packing", False
+        )
+        super(SamplePackingCehrGptDataCollator, self).__init__(*args, **kwargs)
+
+    def __call__(self, examples):
+        current_input_ids = []
+        current_attention_mask = []
+        current_position_ids = []
+        current_value_indicators = []
+        current_values = []
+
+        # Demographics
+        current_person_ids = []
+        current_index_dates = []
+
+        # Binary classification inputs
+        current_ages = []
+        current_labels = []
+
+        for idx, example in enumerate(examples):
+
+            # If the sample length exceeds the model's capacity, truncate this example
+            add_end_token = (
+                len(example["input_ids"]) <= self.max_position_embeddings
+                and self.add_end_token_in_sample_packing
+            )
+
+            if len(example["input_ids"]) > self.max_position_embeddings:
+                example = self.generate_start_end_index(
+                    example, self.max_position_embeddings
+                )
+
+            input_ids = example["input_ids"]
+            # We add [END] [PAD], we want to attend to [END], adding [END] is important for sequence generation.
+            # If the sequence length of the sequence is less than the context window, we add both [END][PAD], otherwise
+            # we only add [PAD] token to the end of the sequence because it's not finished
+            current_input_ids.extend(
+                list(input_ids)
+                + (
+                    [self.tokenizer.end_token_id, self.tokenizer.pad_token_id]
+                    if add_end_token
+                    else [self.tokenizer.pad_token_id]
+                )
+            )
+            current_attention_mask.extend(
+                np.ones_like(input_ids).tolist() + ([1, 0] if add_end_token else [0])
+            )
+            num_tokens_to_pad = 1 + int(add_end_token)
+            current_position_ids.extend(list(range(len(input_ids) + num_tokens_to_pad)))
+            if self.include_values:
+                current_value_indicators.extend(
+                    list(example["value_indicators"]) + [False] * num_tokens_to_pad
+                )
+                current_values.extend(
+                    list(example["values"])
+                    + [self.tokenizer.pad_value_token_id] * num_tokens_to_pad
+                )
+
+            if "person_id" in example:
+                current_person_ids.append(example["person_id"])
+
+            if "index_date" in example:
+                current_index_dates.append(example["index_date"])
+
+            if "age_at_index" in example:
+                current_ages.append(example["age_at_index"])
+
+            if "classifier_label" in example:
+                current_labels.append(example["classifier_label"])
+
+        assert (
+            len(current_input_ids) <= self.max_tokens_per_batch
+        ), f"the total number of tokens in the packed sequence should be less than { self.max_tokens_per_batch}"
+        packed_example = {
+            "input_ids": current_input_ids,
+            "attention_mask": current_attention_mask,
+            "position_ids": current_position_ids,
+        }
+        if self.include_values:
+            packed_example.update({"value_indicators": current_value_indicators})
+            packed_example.update({"values": current_values})
+
+        if current_labels:
+            packed_example.update(
+                {
+                    "person_id": current_person_ids,
+                    "index_date": current_index_dates,
+                    "age_at_index": current_ages,
+                    "classifier_label": current_labels,
+                }
+            )
+
+        return super().__call__([packed_example])