feat: support SFT directly from JSONL without dataset_config.json

- JsonlStore falls back to built-in messages config when no config file found and tokenizer_path is provided
- DatasetFactory.load forwards tokenizer_path to store for SFT/SEQ+jsonl
- assistant turns train, other roles masked, position_ids doc_reset

astrai/dataset/dataset.py

@@ -346,7 +346,15 @@ class DatasetFactory(BaseFactory["BaseDataset"]):
         if processor is not None:
             store.load(load_path, processor=processor, **kwargs)
         else:
-            store.load(load_path, **kwargs)
+            load_kwargs = dict(kwargs)
+            if (
+                tokenizer_path is not None
+                and storage_type == "jsonl"
+                and train_type in ("seq", "sft")
+                and "tokenizer_path" not in load_kwargs
+            ):
+                load_kwargs["tokenizer_path"] = tokenizer_path
+            store.load(load_path, **load_kwargs)
 
         return cls.create(train_type, store=store)
 

astrai/dataset/storage.py

@@ -56,6 +56,7 @@ from typing import Callable, Dict, List, Optional, Tuple, Union
 import torch
 from torch import Tensor
 
+from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
 from astrai.preprocessing.transform import TokenizeTransform
 from astrai.serialization import (
@@ -545,18 +546,29 @@ class JsonlSource:
 
 @StoreFactory.register("jsonl")
 class JsonlStore(Store, Streamable, Recordable):
-    """JSONL reader with two tokenisation modes.
+    """JSONL reader with eager/lazy tokenisation modes.
 
     A JSONL dataset is a ``.jsonl`` file or a directory of ``*.jsonl``
     files plus (optionally) a ``dataset_config.json`` describing the
     tokenization pipeline.
 
-    Two modes, selected at :meth:`load` time:
+    Three ways to supply an eager transform (first match wins):
 
-    - **Eager** (default): applies a :class:`TokenizeTransform` to every
-      record at load time and registers per-key tensors via
-      ``_normalize``.  Both ``fetch`` (stream) and ``fetch_record``
-      (record) work.
+    - **Explicit** (``transform=``): caller-built
+      :class:`TokenizeTransform` applied eagerly.
+    - **Config file**: ``dataset_config.json`` alongside the ``*.jsonl``
+      files — loaded via :meth:`TokenizeTransform.from_config_file`.
+    - **Default messages** (``tokenizer_path=`` given, no config file):
+      a built-in chatml config that tokenises the ``messages`` field,
+      masking every role except ``assistant`` (loss on assistant only).
+      Lets SFT/SEQ train straight from a chat-style JSONL directory
+      without a hand-written config.
+
+    Two tokenisation modes, selected at :meth:`load` time:
+
+    - **Eager** (default): applies the transform to every record at load
+      time and registers per-key tensors via ``_normalize``.  Both
+      ``fetch`` (stream) and ``fetch_record`` (record) work.
     - **Lazy** (``processor=fn`` passed): keeps raw records and defers
       tokenisation to ``fetch_record``.  Only record access works —
       ``len(store)`` returns ``num_records``; stream primitives raise.
@@ -565,6 +577,16 @@ class JsonlStore(Store, Streamable, Recordable):
     CONFIG_NAME = "dataset_config.json"
     segments_are_records = True
 
+    _DEFAULT_MESSAGES_CONFIG = {
+        "version": 1,
+        "input": {
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
+        },
+        "mask": {"system": "mask", "user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "output": {"position_ids_mode": "doc_reset"},
+    }
+
     def __init__(
         self,
         window_size: int = 0,
@@ -587,14 +609,20 @@ class JsonlStore(Store, Streamable, Recordable):
         if transform is None:
             root = Path(path)
             config_path = root / self.CONFIG_NAME if root.is_dir() else None
-            if config_path is None or not config_path.exists():
-                raise FileNotFoundError(
-                    f"JSONL dataset config not found. Expected "
-                    f"{self.CONFIG_NAME} alongside *.jsonl files, pass an "
-                    f"explicit transform, or pass processor= for lazy "
-                    f"on-the-fly tokenisation."
-                )
-            transform = TokenizeTransform.from_config_file(str(config_path))
+            if config_path is not None and config_path.exists():
+                transform = TokenizeTransform.from_config_file(str(config_path))
+            else:
+                tokenizer_path = kwargs.get("tokenizer_path")
+                if not tokenizer_path:
+                    raise FileNotFoundError(
+                        f"JSONL dataset config not found. Expected "
+                        f"{self.CONFIG_NAME} alongside *.jsonl files, pass an "
+                        f"explicit transform, pass processor= for lazy "
+                        f"on-the-fly tokenisation, or pass tokenizer_path= to "
+                        f"use the built-in messages config."
+                    )
+                config = PipelineConfig.from_dict(self._DEFAULT_MESSAGES_CONFIG)
+                transform = TokenizeTransform(config, tokenizer_path)
 
         transformed = transform.apply(records)
         self._normalize(transformed)