improve check for base case

* option to not concatenate during pretraining

* simplify conditional and add doc to config.qmd

docker/Dockerfile-cloud

@@ -20,7 +20,8 @@ RUN apt install --yes --no-install-recommends openssh-server tmux && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
     chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
-    chmod +x /root/cloud-entrypoint.sh
+    chmod +x /root/cloud-entrypoint.sh && \
+    echo 'set-option -g history-limit 5000' >> ~/.tmux.conf
 
 ENTRYPOINT ["/root/cloud-entrypoint.sh"]
 CMD ["sleep", "infinity"]

docs/config.qmd

@@ -244,6 +244,8 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
+# whether to concatenate samples during pretraining
+pretraining_sample_concatenation:
 
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:

scripts/chat_datasets.py

@@ -30,7 +30,7 @@ def parse_dataset(dataset=None, split="train"):
         )
     ds_cfg["field_messages"] = field_messages
 
-    message_fields = features["conversations"][0].keys()
+    message_fields = features[field_messages][0].keys()
     message_field_role = None
     for key in ["from", "role"]:
         if key in message_fields:

src/axolotl/core/trainer_builder.py

@@ -1877,6 +1877,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
     ):
         if training_args.pretraining:
+            if self.cfg.pretraining_sample_concatenation is False:
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/prompt_strategies/chat_template.py

@@ -223,7 +223,7 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
     def tokenize_prompt(self, prompt):
         # Old simple legacy behavior that works reliably.
         if (
-            not self.roles_to_train
+            (not self.roles_to_train or self.roles_to_train == ["assistant"])
             and not self.train_on_eos
             and not self.prompter.message_field_training
             and not self.prompter.message_field_training_detail

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -706,6 +706,12 @@ class AxolotlInputConfig(
     pad_to_sequence_len: Optional[bool] = None
     curriculum_sampling: Optional[bool] = None
     multipack_real_batches: Optional[bool] = None
+    pretraining_sample_concatenation: Optional[bool] = Field(
+        default=None,
+        json_schema_extra={
+            "description": "whether to soft pack/concatenate samples during pretraining",
+        },
+    )
 
     batch_flattening: Optional[Union[Literal["auto"], bool]] = None
 

src/axolotl/utils/data/pretraining.py

@@ -22,6 +22,7 @@ def encode_pretraining(
     max_tokens: int,
     examples: Dict[str, List],
     text_column: str = "text",
+    concatenate: bool = True,
 ) -> Dict[str, List]:
     res = tokenizer(
         examples[text_column],
@@ -33,6 +34,13 @@ def encode_pretraining(
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
     targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
+    if not concatenate:
+        return {
+            "input_ids": [seq.tolist() for seq in input_ids],
+            "labels": [seq.tolist() for seq in targets],
+            "attention_mask": [seq.tolist() for seq in attention_mask],
+        }
+
     new_input_ids = []
     new_labels = []
     new_attention_mask = []
@@ -204,6 +212,7 @@ def wrap_pretraining_dataset(
             tokenizer,
             max_tokens,
             text_column=cfg.pretraining_dataset[0].text_column or "text",
+            concatenate=cfg.pretraining_sample_concatenation is True,
         )
 
     if cfg.shuffle_merged_datasets: