review comments, tests, etc.

cli/commands/train.py

@@ -19,6 +19,9 @@ def train(
     hf_token: Optional[str] = typer.Option(
         None, "--hf-token", envvar="HF_TOKEN", help="Hugging Face token if needed."
     ),
+    wandb_token: Optional[str] = typer.Option(
+        None, "--wandb-token", envvar="WANDB_API_KEY", help="Weights & Biases API key."
+    ),
     dry_run: bool = typer.Option(
         False,
         "--dry-run",
@@ -35,6 +38,10 @@ def train(
 
     cfg.apply_overrides(**config_overrides)
 
+    # CLI/env tokens take precedence over config
+    hf_token = hf_token or cfg.logging.hf_token
+    wandb_token = wandb_token or cfg.logging.wandb_token
+
     if dry_run:
         import yaml
         data = cfg.model_dump()
@@ -52,13 +59,18 @@ def train(
         )
         raise typer.Exit(code=2)
 
+    from pathlib import Path as PathlibPath
     from backend.trainer import UnslothTrainer
     from backend.model_config import ModelConfig
 
     trainer = UnslothTrainer()
 
+    # Check if the model path is a LoRA adapter (has adapter_config.json)
+    model_path = PathlibPath(cfg.model) if cfg.model else None
+    model_is_lora = model_path and model_path.is_dir() and (model_path / "adapter_config.json").exists()
+
     model_config = ModelConfig.from_ui_selection(
-        dropdown_value=cfg.model, search_value=None, hf_token=hf_token, is_lora=False
+        dropdown_value=cfg.model, search_value=None, hf_token=hf_token, is_lora=model_is_lora
     )
     if not model_config:
         typer.echo("Could not resolve model config", err=True)
@@ -67,6 +79,14 @@ def train(
     is_vision = model_config.is_vision
     use_lora = cfg.training.training_type.lower() == "lora"
 
+    if model_is_lora and not use_lora:
+        typer.echo(
+            "Error: Cannot do full finetuning on a LoRA adapter. "
+            "Use --training-type lora or provide a base model.",
+            err=True,
+        )
+        raise typer.Exit(code=2)
+
     if not trainer.load_model(
         model_name=model_config.identifier,
         max_seq_length=cfg.training.max_seq_length,
@@ -89,7 +109,9 @@ def train(
         typer.echo("Dataset load failed", err=True)
         raise typer.Exit(code=1)
 
-    started = trainer.start_training(dataset=ds, **cfg.training_kwargs())
+    training_kwargs = cfg.training_kwargs()
+    training_kwargs["wandb_token"] = wandb_token  # CLI/env takes precedence
+    started = trainer.start_training(dataset=ds, **training_kwargs)
 
     if not started:
         typer.echo("Training failed to start", err=True)

cli/config.py

@@ -1,5 +1,5 @@
 from pathlib import Path
-from typing import Optional, List
+from typing import Literal, Optional, List
 
 import yaml
 from pydantic import BaseModel, Field
@@ -8,11 +8,11 @@ from pydantic import BaseModel, Field
 class DataConfig(BaseModel):
     dataset: Optional[str] = None
     local_dataset: Optional[List[str]] = None
-    format_type: str = "auto"
+    format_type: Literal["auto", "alpaca", "chatml", "sharegpt"] = "auto"
 
 
 class TrainingConfig(BaseModel):
-    training_type: str = "lora"
+    training_type: Literal["lora", "full"] = "lora"
     max_seq_length: int = 2048
     load_in_4bit: bool = True
     output_dir: Path = Path("./outputs")
@@ -27,7 +27,7 @@ class TrainingConfig(BaseModel):
     random_seed: int = 3407
     packing: bool = False
     train_on_completions: bool = False
-    gradient_checkpointing: bool = True
+    gradient_checkpointing: Literal["unsloth", "true", "none"] = "unsloth"
 
 
 class LoraConfig(BaseModel):
@@ -53,6 +53,7 @@ class LoggingConfig(BaseModel):
     wandb_token: Optional[str] = None
     enable_tensorboard: bool = False
     tensorboard_dir: str = "runs"
+    hf_token: Optional[str] = None
 
 
 class Config(BaseModel):

cli/options.py

@@ -82,11 +82,15 @@ def add_options_from_config(config_class: type[BaseModel]) -> Callable:
     def decorator(func: Callable) -> Callable:
         sig = inspect.signature(func)
         original_params = list(sig.parameters.values())
+        original_param_names = {p.name for p in original_params}
 
         # Build new parameters: config fields first, then original params
         new_params = []
 
         for field_name, field_info in fields:
+            # Skip fields already defined in function signature (e.g., with envvar)
+            if field_name in original_param_names:
+                continue
             annotation = field_info.annotation
             if _is_list_type(annotation):
                 continue