fix: use AutoModel for audio, is_relative_to path check, early librosa import

- Use AutoModel instead of AutoModelForCausalLM for audio-language models
  (Qwen2-Audio, Whisper don't work with causal LM auto class)
- Use Path.is_relative_to() for path traversal check (symlink-safe, Python 3.9+)
- Fail fast with helpful error if librosa not installed before dataset processing

soup_cli/data/loader.py

@@ -195,7 +195,7 @@ def _validate_audio_files(data: list[dict], audio_dir: Path) -> list[dict]:
             audio_path = audio_dir / audio_path
         # Path traversal protection: resolved path must stay under audio_dir
         resolved = audio_path.resolve()
-        if not str(resolved).startswith(str(resolved_base)):
+        if not resolved.is_relative_to(resolved_base):
             traversal += 1
             continue
         valid.append({**row, "audio": str(resolved)})

soup_cli/trainer/sft.py

@@ -424,7 +424,7 @@ class SFTTrainerWrapper:
         """Load audio-language model via transformers (Qwen2-Audio, Whisper, etc.)."""
         from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
         from rich.panel import Panel as RichPanel
-        from transformers import AutoModelForCausalLM, AutoProcessor, BitsAndBytesConfig
+        from transformers import AutoModel, AutoProcessor, BitsAndBytesConfig
 
         console.print(
             RichPanel(
@@ -461,7 +461,9 @@ class SFTTrainerWrapper:
         if bnb_config:
             model_kwargs["quantization_config"] = bnb_config
 
-        self.model = AutoModelForCausalLM.from_pretrained(cfg.base, **model_kwargs)
+        # Use AutoModel for audio models — AutoModelForCausalLM doesn't handle
+        # audio-language architectures (Qwen2-Audio, Whisper, etc.)
+        self.model = AutoModel.from_pretrained(cfg.base, **model_kwargs)
 
         if tcfg.quantization in ("4bit", "8bit"):
             self.model = prepare_model_for_kbit_training(self.model)
@@ -485,6 +487,14 @@ class SFTTrainerWrapper:
         """Prepare dataset for audio fine-tuning with audio loading."""
         from datasets import Dataset
 
+        try:
+            import librosa  # noqa: F401
+        except ImportError:
+            raise ImportError(
+                "librosa is required for audio training. "
+                "Install with: pip install 'soup-cli[audio]'"
+            )
+
         def load_and_format_audio(example):
             import librosa