Exposes API for processing pretraining data

README.md

@@ -25,6 +25,7 @@ The library now supports reasoning traces through the `reasoning_content` field
 - [Using the library](#using-the-library)
 - [Data format](#data-format)
   - [Reasoning content support](#reasoning-content-support-1)
+- [Continual pretraining mode](#continual-pretraining-mode)
 - [Documentation](#documentation)
 - [Learning about the training arguments](#learning-about-training-arguments)
   - [`TrainingArgs`](#trainingargs)
@@ -122,6 +123,46 @@ The library now supports an optional `reasoning_content` field in addition to th
 }
 ```
 
+## Continual pretraining mode
+
+In addition to instruction tuning, the library can run document-style continual pretraining on raw text corpora.
+Enable this by supplying a block size when invoking `main_ds.py`:
+
+```bash
+torchrun main_ds.py \
+  --model_name_or_path mistralai/Mistral-7B-v0.1 \
+  --data_path /data/documents.jsonl \
+  --ckpt_output_dir ./checkpoints \
+  --effective_batch_size 128 \
+  --max_batch_len 60000 \
+  --block-size 8192 \
+  --document-column-name text  # optional, defaults to "document"
+```
+
+- `--block-size` (required) toggles continual pretraining and controls how many tokens are packed into each block.
+- `--document-column-name` (optional) specifies which JSONL field contains the raw document text.
+
+The same options are available programmatically via `TrainingArgs.pretraining_config`:
+
+```python
+from instructlab.training import TrainingArgs, PretrainingConfig
+
+train_args = TrainingArgs(
+    model_name_or_path="mistralai/Mistral-7B-v0.1",
+    data_path="documents.jsonl",
+    ckpt_output_dir="./checkpoints",
+    max_seq_len=4096,
+    max_batch_len=40000,
+    effective_batch_size=128,
+    pretraining_config=PretrainingConfig(
+        block_size=2048,
+        document_column_name="text",  # optional
+    ),
+)
+```
+
+When a pretraining config is provided, `process_documents_for_pretraining()` is invoked under the hood to tokenize raw documents before training.
+
 **Standard message structure:**
 
 ```json
@@ -139,7 +180,7 @@ The library now supports an optional `reasoning_content` field in addition to th
 }
 ```
 
-#### Important Notes
+### Important Notes
 
 1. **Automatic reasoning content processing**: If `reasoning_content` exists in a message, it will always be processed and unmasked as long as the message role is targeted for unmasking. This ensures that reasoning traces are properly included in the training data.
 

src/instructlab/training/__init__.py

@@ -10,6 +10,7 @@
     "FSDPOptions",
     "ShardingStrategies",
     "DistributedBackend",
+    "PretrainingConfig",
 )
 
 # First Party
@@ -23,6 +24,7 @@
     DistributedBackend,
     FSDPOptions,
     LoraOptions,
+    PretrainingConfig,
     QuantizeDataType,
     ShardingStrategies,
     TorchrunArgs,

src/instructlab/training/accelerator.py

@@ -63,14 +63,17 @@ def __init__(
         self.lr_scheduler = None
         if self.distributed_framework == DistributedBackend.DEEPSPEED:
             # Standard
+            cpu_offload_optimizer_ratio = (
+                self.deepspeed_cpu_offload_optimizer_ratio or 0.0
+            )
             accel_args = {
                 "deepspeed_plugin": self.get_ds_plugin(
                     world_size=torch.distributed.get_world_size(),
                     samples_per_gpu=samples_per_gpu,
                     grad_accum=grad_accum,
                     opts=DeepSpeedOptions(
                         cpu_offload_optimizer=deepspeed_cpu_offload_optimizer,
-                        cpu_offload_optimizer_ratio=self.deepspeed_cpu_offload_optimizer_ratio,
+                        cpu_offload_optimizer_ratio=cpu_offload_optimizer_ratio,
                         cpu_offload_optimizer_pin_memory=self.deepspeed_cpu_offload_optimizer_pin_memory,
                         save_samples=save_samples,
                     ),

src/instructlab/training/config.py

@@ -67,6 +67,20 @@ class DataProcessArgs(BaseModel):
     model_config = ConfigDict(protected_namespaces=())
 
 
+class PretrainingConfig(BaseModel):
+    """
+    Configuration for pretraining mode.
+    """
+
+    block_size: int = Field(
+        description="Size of each block in tokens for pretraining datasets."
+    )
+    document_column_name: str = Field(
+        default="document",
+        description="Name of the column containing raw documents for pretraining.",
+    )
+
+
 # public API
 class TorchrunArgs(BaseModel):
     """
@@ -266,6 +280,14 @@ class TrainingArgs(BaseModel):
     # "last_epoch". This works alongside the '--checkpoint_at_epoch' flag.
     keep_last_checkpoint_only: Optional[bool] = False
 
+    pretraining_config: Optional[PretrainingConfig] = Field(
+        default=None,
+        description=(
+            "Pretraining configuration. When provided, enables block-based sampling "
+            "for raw document pretraining datasets."
+        ),
+    )
+
     # TODO(osilkin):
     #   we are only exposing this here because `run_training` today is implicitly coupled
     #   with `process_data`. Since we don't have a specific field for data processing arguments,

src/instructlab/training/data_process.py

@@ -412,7 +412,10 @@ def process_messages_into_input_ids_with_chat_template(args: DataProcessArgs):
     logger.info("Tokenizing the dataset with %s tokenizer...", args.model_path)
     data_with_input_ids = data.map(
         lambda x: {
-            "input_ids": tokenizer.apply_chat_template(x["messages"], tokenize=True),
+            # newer versions of transformers have `return_dict=True` by default
+            "input_ids": tokenizer.apply_chat_template(
+                x["messages"], tokenize=True, return_dict=False
+            ),
             "unmask": bool(x["unmask"]) if "unmask" in x else False,
         },
         num_proc=NUM_PROC,
@@ -687,7 +690,8 @@ def unmask_messages(
             if regions:
                 message_regions_map[idx] = regions
 
-    input_ids = tokenizer.apply_chat_template(msgs_with_unmasking)
+    # newer versions of transformers have `return_dict=True` by default
+    input_ids = tokenizer.apply_chat_template(msgs_with_unmasking, return_dict=False)
 
     # Get token IDs for all unmask tokens
     unmask_begin_token_id = tokenizer.encode(
@@ -1133,6 +1137,109 @@ def process_messages_into_input_ids(
     save_dataset(final_dataset, data_output_path, num_cpu_procs)
 
 
+def process_documents_for_pretraining(
+    data_path: str,
+    data_output_path: str,
+    model_path: str,
+    num_cpu_procs: int,
+    document_column_name: str = "document",
+) -> None:
+    """
+    Process raw documents for pretraining by tokenizing without chunking.
+
+    Outputs one JSONL record per document with only input_ids (no labels).
+    Blocking/chunking happens later during training.
+
+    Pattern: Each document → [BOS][tokens][EOS]
+
+    Args:
+        data_path: Path to input JSONL with {"document": "text"} format
+        data_output_path: Directory for processed data output
+        model_path: Path to model/tokenizer
+        num_cpu_procs: Number of parallel processes
+        document_column_name: Name of the column containing the documents
+    """
+    ensure_can_write_to_directory(data_output_path)
+
+    # Load and validate dataset
+    try:
+        data = load_dataset("json", data_files=data_path, split="train")
+    except Exception as e:
+        raise ValueError(
+            "Malformed or missing data, please ensure your dataset is correctly formatted"
+        ) from e
+
+    if data.num_rows == 0:
+        raise ValueError("The provided dataset is empty")
+
+    if document_column_name not in data.column_names:
+        raise ValueError(
+            f"Pretraining data must have '{document_column_name}' field. Found: {data.column_names}"
+        )
+
+    logger.info("Loading tokenizer from %s", model_path)
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+    if tokenizer.eos_token_id is None:
+        raise ValueError("Tokenizer must have an EOS token defined for pretraining")
+
+    logger.info("Tokenizing %d documents for pretraining...", data.num_rows)
+
+    # Tokenize each document: encode() adds BOS, then append EOS
+    def tokenize_document(sample):
+        input_ids = tokenizer.encode(
+            sample[document_column_name], add_special_tokens=True
+        )
+
+        # ensures eos token is present without double-adding it.
+        if input_ids[-1] != tokenizer.eos_token_id:
+            input_ids.append(tokenizer.eos_token_id)
+
+        return {
+            "input_ids": input_ids,
+            "len": len(input_ids),
+        }
+
+    # Filter out empty documents before tokenization
+    def filter_empty_documents(batch):
+        return [bool(doc) for doc in batch[document_column_name]]
+
+    filtered_data = data.filter(
+        filter_empty_documents,
+        batched=True,
+        num_proc=num_cpu_procs,
+        desc="Filtering empty documents",
+    )
+
+    dropped_count = data.num_rows - filtered_data.num_rows
+    if dropped_count > 0:
+        logger.info(f"Dropped {dropped_count:,} empty documents")
+    tokenized_data = filtered_data.map(
+        tokenize_document,
+        num_proc=num_cpu_procs,
+        desc="Tokenizing documents",
+        remove_columns=filtered_data.column_names,
+    )
+
+    # Calculate statistics
+    total_tokens = sum(tokenized_data["len"])
+    avg_tokens = total_tokens / len(tokenized_data)
+    logger.info(f"Processed {len(tokenized_data):,} documents")
+    logger.info(f"Total tokens: {total_tokens:,}")
+    logger.info(f"Average tokens per document: {avg_tokens:.1f}")
+
+    # Save to JSONL (one record per document)
+    os.makedirs(data_output_path, exist_ok=True)
+    output_file = Path(data_output_path) / "data.jsonl"
+
+    tokenized_data.to_json(
+        output_file, num_proc=num_cpu_procs, lines=True, orient="records"
+    )
+
+    logger.info(f"Saved tokenized documents to {output_file}")
+    logger.info("Note: Blocking into fixed-size chunks will happen during training")
+
+
 def ensure_can_write_to_directory(output_dir: str) -> None:
     """
     Ensure that we can write to the output directory.