Fix KD experiment path collision between FP32 and BitNet students

- Use version-specific prefix (std_kd vs bit_kd) for KD experiments
- Pass version parameter from train_kd.py to get_experiment_dir()
- Resolves collision where Phase 2 (FP32+KD) and Phase 2.9 (BitNet+KD)
  were writing to the same bit_kd_s* directories
- Existing Phase 2 experiments need manual rename: bit_kd_* → std_kd_*

Author: dariocazzani

experiments/paths.py

@@ -85,7 +85,9 @@ def get_experiment_dir(
         if lr != TRAINING_DEFAULTS.lr:
             parts.append(f"lr{lr:g}")
     else:
-        parts.append("bit_kd")
+        # Use version-specific prefix for KD experiments (std_kd vs bit_kd)
+        kd_prefix = f"{version}_kd"
+        parts.append(kd_prefix)
         if ablation != TRAINING_DEFAULTS.ablation:
             parts.append(ablation)
         if optimizer != TRAINING_DEFAULTS.optimizer:

experiments/train_kd.py

@@ -281,11 +281,13 @@ def main() -> None:
 
     # Auto-generate output dir if not specified
     if not args.output_dir:
+        version_str = "std" if args.student_is_fp32 else "bit"
         experiment_dir = get_experiment_dir(
             ExperimentType.KD,
             args.dataset,
             args.model,
             args.seed,
+            version=version_str,
             ablation=args.ablation,
             lr=args.lr,
             optimizer=args.optimizer,