exp: SAK_LR override; 4e-4 diverges to NaN at 6 epochs

job-0.5b-exp.py

@@ -137,10 +137,12 @@ if MODE == "lora-masked":
         r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM", use_rslora=True,
         target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]))
     model.print_trainable_parameters()
-    lr = 4e-4
+    # 4e-4 (sqrt-scaled for batch 64) is marginally unstable: it survives 3 epochs
+    # but diverges to NaN by 6. Overridable so longer runs can drop it.
+    lr = float(os.environ.get("SAK_LR", "4e-4"))
 else:
     # Full fine-tune needs a far lower LR than LoRA; 4e-4 would destroy the base.
-    lr = 2e-5
+    lr = float(os.environ.get("SAK_LR", "2e-5"))
     print("full fine-tune: all", sum(p.numel() for p in model.parameters()), "params trainable")
 
 args = SFTConfig(output_dir=f"out-{MODE}-{TAG}", num_train_epochs=EPOCHS, seed=SEED,