more fixes

fix loader default

src/axolotl/common/datasets.py

@@ -11,7 +11,7 @@ from datasets import Dataset
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
 from axolotl.utils.data import prepare_dataset
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
@@ -109,9 +109,9 @@ def load_preference_datasets(
     cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
 ) -> TrainDatasetMeta:
     """
-    Loads one or more training or evaluation datasets for DPO training, calling
+    Loads one or more training or evaluation datasets for RL training using paired
-    `axolotl.utils.data.rl.load_prepare_dpo_datasets`. Optionally, logs out debug
+    preference data, calling `axolotl.utils.data.rl.load_prepare_preference_datasets`.
-    information.
+    Optionally, logs out debug information.
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
@@ -121,7 +121,7 @@ def load_preference_datasets(
         Dataclass with fields for training and evaluation datasets and the computed
         `total_num_steps`.
     """
-    train_dataset, eval_dataset = load_prepare_dpo_datasets(cfg)
+    train_dataset, eval_dataset = load_prepare_preference_datasets(cfg)
     total_num_steps = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )

src/axolotl/integrations/kd/chat_template.py

@@ -53,6 +53,10 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         )
 
     def transform_logprobs(self, sample):
+        """
+        Transform logprobs to target format for KD training
+        """
+
         logprobs = sample.pop(self.logprobs_field)
         target_seq_len = len(logprobs)
         input_seq_len = len(sample["input_ids"])
@@ -62,16 +66,33 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         target_token_ids = []
         target_mask = []
 
+        if input_padding_len < 0:
+            # logprobs is longer than target_seq_len,
+            # so we need to slice from the left/beginning of logprobs
+            logprobs = logprobs[:-input_seq_len]
+            input_padding_len = 0
+            # target_seq_len = input_seq_len
+
         # fill with -inf for padding_len tokens for top_k tokens
         # extend target_logprobs with a padding_len x top_k 2D list filled with -inf
-        for _ in range(1, input_padding_len):  # start at 1 since this is causal
+
+        # for causal models, if we start the range at 1, then we don't need to shift in the trainer
+        # otherwise, we need to shift in the trainer
+        shift = 0
+        for _ in range(shift, input_padding_len):
             target_logprobs.append([-float("inf")] * top_k)
             target_token_ids.append(list(range(top_k)))
             target_mask.append([0] * top_k)
 
-        for _ in range(target_seq_len):
+        # for _ in range(target_seq_len):
-            # TODO also check against sample["labels"]
+        #     # TODO also check against sample["labels"]
-            target_mask.append([1] * top_k)
+        #     target_mask.append([1] * top_k)
+
+        for position in range(input_padding_len, input_seq_len):
+            if sample["labels"][position] == -100:
+                target_mask.append([0] * top_k)
+            else:
+                target_mask.append([1] * top_k)
 
         for _, token_pos_logprobs in enumerate(logprobs):
             # Initialize collections for logprobs and token_ids
@@ -120,10 +141,11 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             target_logprobs.append(position_logprobs_scaled)
             target_token_ids.append(position_token_ids)
 
-        # since we started at index 1 for causal, we need one more padding token
+        if shift == 1:
-        target_logprobs.append([-float("inf")] * top_k)
+            # since we started at index 1 for causal, we need one more padding token
-        target_token_ids.append(list(range(top_k)))
+            target_logprobs.append([-float("inf")] * top_k)
-        target_mask.append([0] * top_k)
+            target_token_ids.append(list(range(top_k)))
+            target_mask.append([0] * top_k)
 
         # Update sample with transformed logprobs
         sample["target_logprobs"] = target_logprobs

src/axolotl/integrations/kd/trainer.py

@@ -45,7 +45,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
         inputs,
         return_outputs=False,
         num_items_in_batch=None,
-        shift_targets=False,
+        shift_targets=True,
     ):
         """
         How the loss is computed by Trainer. By default, all models return the loss in the first element.

src/axolotl/utils/data/__init__.py

@@ -5,7 +5,7 @@ from axolotl.utils.data.pretraining import (  # noqa: F401
     encode_pretraining,
     wrap_pretraining_dataset,
 )
-from axolotl.utils.data.rl import load_prepare_dpo_datasets  # noqa: F401
+from axolotl.utils.data.rl import load_prepare_preference_datasets  # noqa: F401
 from axolotl.utils.data.sft import (  # noqa: F401
     get_dataset_wrapper,
     load_prepare_datasets,

src/axolotl/utils/data/pretraining.py

@@ -18,10 +18,13 @@ LOG = logging.getLogger("axolotl")
 
 
 def encode_pretraining(
-    tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: Dict[str, List]
+    tokenizer: PreTrainedTokenizerBase,
+    max_tokens: int,
+    examples: Dict[str, List],
+    text_column: str = "text",
 ) -> Dict[str, List]:
     res = tokenizer(
-        examples["text"],
+        examples[text_column],
         truncation=True,
         max_length=max_tokens - 2,
         add_special_tokens=True,
@@ -196,7 +199,12 @@ def wrap_pretraining_dataset(
         # set this to 1 so downstream data_loader doesn't try to increase the batch again
         cfg.micro_batch_size = 1
     else:
-        encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
+        encode = functools.partial(
+            encode_pretraining,
+            tokenizer,
+            max_tokens,
+            text_column=cfg.pretraining_dataset[0].text_column or "text",
+        )
 
     if cfg.shuffle_merged_datasets:
         dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)

src/axolotl/utils/data/rl.py

@@ -115,7 +115,7 @@ def drop_long_rl_seq(
     raise ValueError("Unknown RL type")
 
 
-def load_prepare_dpo_datasets(cfg):
+def load_prepare_preference_datasets(cfg):
     def load_split(dataset_cfgs, _cfg):
         split_datasets: List[Any] = []
         for i, ds_cfg in enumerate(dataset_cfgs):

src/axolotl/utils/models.py

@@ -1057,7 +1057,7 @@ class ModelLoader:
         )
         if (
             hasattr(self.model, "get_input_embeddings")
-            and self.model.get_input_embeddings().num_embeddings < embeddings_len
+            and self.model.get_input_embeddings().num_embeddings != embeddings_len
         ):
             resize_kwargs = {}
             if self.cfg.mean_resizing_embeddings is not None:

src/axolotl/utils/trainer.py

@@ -279,6 +279,7 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         drop_long_kwargs["desc"] = "Dropping Long Sequences"
     train_dataset = train_dataset.filter(
         drop_long,
+        batched=True,
         **filter_map_kwargs,
         **drop_long_kwargs,
     )
@@ -310,8 +311,7 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         """
         labels = sample["labels"]
         if not labels:
-            # Edge case: if labels is empty, decide if you want to keep or drop
+            return True
-            return True  # or False
 
         # Check if single example or batch
         # If first element is an int, we assume a single example

tests/e2e/test_llama_pretrain.py

@@ -4,7 +4,8 @@ E2E tests for llama pretrain
 
 import logging
 import os
-import unittest
+
+import pytest
 
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
@@ -12,19 +13,22 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import check_model_output_exists
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestPretrainLlama(unittest.TestCase):
+class TestPretrainLlama:
     """
     Test case for Llama models w pretraining
     """
 
-    @with_temp_dir
+    @pytest.mark.parametrize(
-    def test_pretrain_w_sample_packing(self, temp_dir):
+        "sample_packing",
+        [True, False],
+    )
+    def test_pretrain(self, temp_dir, sample_packing):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -32,7 +36,7 @@ class TestPretrainLlama(unittest.TestCase):
                 "tokenizer_type": "LlamaTokenizer",
                 "flash_attention": True,
                 "sequence_len": 1024,
-                "sample_packing": True,
+                "sample_packing": sample_packing,
                 "special_tokens": {
                     "unk_token": "<unk>",
                     "bos_token": "<s>",

tests/test_datasets.py

@@ -17,7 +17,7 @@ from huggingface_hub import snapshot_download
 from transformers import AutoTokenizer
 
 from axolotl.utils.data import load_tokenized_prepared_datasets
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 
 
@@ -280,7 +280,7 @@ class TestDatasetPreparation(unittest.TestCase):
             }
         )
 
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
+        train_dataset, _ = load_prepare_preference_datasets(cfg)
 
         assert len(train_dataset) == 1800
         assert "conversation" in train_dataset.features
@@ -329,7 +329,7 @@ class TestDatasetPreparation(unittest.TestCase):
             }
         )
 
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
+        train_dataset, _ = load_prepare_preference_datasets(cfg)
 
         assert len(train_dataset) == 1800
         assert "conversation" in train_dataset.features

tests/test_exact_deduplication.py

@@ -12,7 +12,7 @@ from datasets import Dataset
 from transformers import AutoTokenizer
 
 from axolotl.utils.data import prepare_dataset
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.data.utils import deduplicate_and_log_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
@@ -236,7 +236,7 @@ class TestDeduplicateRLDataset(unittest.TestCase):
         """Verify that loading with deduplication removes duplicates."""
 
         # Load the dataset using the deduplication setting
-        train_dataset, _ = load_prepare_dpo_datasets(self.cfg)
+        train_dataset, _ = load_prepare_preference_datasets(self.cfg)
 
         # Verify that the dataset has been deduplicated
         assert len(train_dataset) == 1800, "Dataset was not properly deduplicated"
@@ -245,7 +245,7 @@ class TestDeduplicateRLDataset(unittest.TestCase):
         """Verify that loading without deduplication retains duplicates."""
         self.cfg.dataset_exact_deduplication = False
         # Load the dataset without deduplication
-        train_dataset, _ = load_prepare_dpo_datasets(self.cfg)
+        train_dataset, _ = load_prepare_preference_datasets(self.cfg)
 
         # Verify that the dataset retains duplicates
         assert (