QLoRA training kernel for Qwen2.5-7B-Instruct on a T4

train_qwen.py

@@ -21,8 +21,19 @@ MODEL = "Qwen/Qwen2.5-7B-Instruct"
 HF_TOKEN = os.environ.get("HF_TOKEN", "")
 assert HF_TOKEN, "set HF_TOKEN env var (never hardcode tokens in code)"
 OUT_ADAPTER = "/kaggle/working/adapter"
-DATA = "/kaggle/working/synthetic_train.jsonl"
-VAL = "/kaggle/working/synthetic_val.jsonl"
+
+# ---------- locate data: Kaggle dataset mount first, then /kaggle/working ----------
+import glob
+DATA = VAL = None
+for cand in glob.glob("/kaggle/input/*/synthetic_train.jsonl"):
+    DATA = cand
+    VAL = cand.replace("synthetic_train.jsonl", "synthetic_val.jsonl")
+    break
+if DATA is None:
+    DATA, VAL = "/kaggle/working/synthetic_train.jsonl", "/kaggle/working/synthetic_val.jsonl"
+assert os.path.exists(DATA), "synthetic_train.jsonl not found"
+print("data:", DATA, os.path.getsize(DATA), "bytes")
+
 
 assert torch.cuda.is_available(), "No GPU — set accelerator to GPU T4 x1"
 print("torch", torch.__version__, "|", torch.cuda.get_device_name(0))