Update runners/KAGGLE_TRAIN_SARUS_SFT.txt with user enhancements

runners/KAGGLE_TRAIN_SARUS_SFT.txt

@@ -1,67 +1,68 @@
-# ==============================================================================
-# 🚀 ViuAI Sarus-500M — STEP 2: Kaggle Master SFT v1 Training Engine
-# ==============================================================================
-# Run this entire block in a single Kaggle Notebook cell (Accelerator: GPU T4 x 2 or P100)
-# Auto-syncs:
-#   • Base weights:     ViuAI/ViuAI-500M (checkpoints/ckpt_latest.pt)
-#   • Tokenized data:   ViuAI/viuai-500m-sft-tokenized (sft_v1)
-#   • Training Engine:  code/train_sft.py
-# ==============================================================================
-
-import os, sys, shutil, subprocess
-
-# 1. Environment & Auth Token
-HF_TOKEN = os.environ.get("HF_TOKEN") or ("".join(["hf_", "ssyCVhuny", "XxjGdqKp", "VLPpkmWK", "FrrMOIFbg"]))
-os.environ["HF_TOKEN"] = HF_TOKEN
-os.environ["HF_HUB_DISABLE_PROGRESS_BARS"] = "1"
-os.environ["TOKENIZERS_PARALLELISM"] = "false"
-os.environ["PYTHONUNBUFFERED"] = "1"
-
-# 2. Dependencies
-!pip install -q torch tokenizers huggingface_hub numpy
-
-from huggingface_hub import hf_hub_download
-from huggingface_hub.utils import disable_progress_bars
-disable_progress_bars()
-
-# 3. Setup Workspace Folders
-os.makedirs("code", exist_ok=True)
-os.makedirs("tokenizer", exist_ok=True)
-os.makedirs("checkpoints", exist_ok=True)
-os.makedirs("tokenized_data/sft_v1", exist_ok=True)
-
-# 4. Fetch Architecture & Training Script
-print("📦 Fetching Sarus Model Architecture & SFT Training Engine from Hugging Face...")
-code_files = [
-    ("code/model.py", "code/model.py"),
-    ("code/config.py", "code/config.py"),
-    ("code/train_sft.py", "code/train_sft.py"),
-    ("tokenizer/tokenizer.json", "tokenizer/tokenizer.json"),
-]
-
-for src, dst in code_files:
-    p = hf_hub_download(repo_id="ViuAI/ViuAI-500M", filename=src, token=HF_TOKEN)
-    shutil.copy(p, dst)
-print("   ✅ Architecture and training engine ready.")
-
-# 5. Fetch Base Checkpoint (ckpt_latest.pt ~5.9GB) - Skipped if already present
-base_ckpt = "checkpoints/ckpt_latest.pt"
-if not os.path.exists(base_ckpt):
-    print("📥 Downloading base checkpoint ckpt_latest.pt from ViuAI/ViuAI-500M...")
-    dl_ckpt = hf_hub_download(repo_id="ViuAI/ViuAI-500M", filename=base_ckpt, token=HF_TOKEN)
-    if dl_ckpt != base_ckpt:
-        shutil.copy(dl_ckpt, base_ckpt)
-print("   ✅ Base checkpoint ready.")
-
-# 6. Execute Master SFT v1 Training Engine
-print("\n🔥 Launching Sarus SFT v1 Master Training on GPU...")
-!python -u code/train_sft.py \
-    --version v1 \
-    --epochs 3 \
-    --batch_size 4 \
-    --grad_accum 32 \
-    --max_lr 3.2e-5 \
-    --min_lr 2.0e-6 \
-    --neftune_alpha 5.0 \
-    --eval_interval 500 \
-    --push_to_hf
+# ==============================================================================
+# 🚀 ViuAI Sarus-500M — STEP 2: Kaggle Master SFT v1 Training Engine
+# ==============================================================================
+# Run this entire block in a single Kaggle Notebook cell (Accelerator: GPU T4 x 2 or P100)
+# Auto-syncs:
+#   • Base weights:     ViuAI/ViuAI-500M (checkpoints/ckpt_latest.pt)
+#   • Tokenized data:   ViuAI/viuai-500m-sft-tokenized (sft_v1)
+#   • Training Engine:  code/train_sft.py
+# ==============================================================================
+
+import os, sys, shutil, subprocess
+
+# 1. Environment & Auth Token
+HF_TOKEN = os.environ.get("HF_TOKEN")
+if HF_TOKEN:
+    os.environ["HF_TOKEN"] = HF_TOKEN
+os.environ["HF_HUB_DISABLE_PROGRESS_BARS"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+os.environ["PYTHONUNBUFFERED"] = "1"
+
+# 2. Dependencies
+!pip install -q torch tokenizers huggingface_hub numpy
+
+from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import disable_progress_bars
+disable_progress_bars()
+
+# 3. Setup Workspace Folders
+os.makedirs("code", exist_ok=True)
+os.makedirs("tokenizer", exist_ok=True)
+os.makedirs("checkpoints", exist_ok=True)
+os.makedirs("tokenized_data/sft_v1", exist_ok=True)
+
+# 4. Fetch Architecture & Training Script
+print("📦 Fetching Sarus Model Architecture & SFT Training Engine from Hugging Face...")
+code_files = [
+    ("code/model.py", "code/model.py"),
+    ("code/config.py", "code/config.py"),
+    ("code/train_sft.py", "code/train_sft.py"),
+    ("tokenizer/tokenizer.json", "tokenizer/tokenizer.json"),
+]
+
+for src, dst in code_files:
+    p = hf_hub_download(repo_id="ViuAI/ViuAI-500M", filename=src, token=HF_TOKEN)
+    shutil.copy(p, dst)
+print("   ✅ Architecture and training engine ready.")
+
+# 5. Fetch Base Checkpoint (ckpt_latest.pt ~5.9GB) - Skipped if already present
+base_ckpt = "checkpoints/ckpt_latest.pt"
+if not os.path.exists(base_ckpt):
+    print("📥 Downloading base checkpoint ckpt_latest.pt from ViuAI/ViuAI-500M...")
+    dl_ckpt = hf_hub_download(repo_id="ViuAI/ViuAI-500M", filename=base_ckpt, token=HF_TOKEN)
+    if dl_ckpt != base_ckpt:
+        shutil.copy(dl_ckpt, base_ckpt)
+print("   ✅ Base checkpoint ready.")
+
+# 6. Execute Master SFT v1 Training Engine
+print("\n🔥 Launching Sarus SFT v1 Master Training on GPU...")
+!python -u code/train_sft.py \
+    --version v1 \
+    --epochs 3 \
+    --batch_size 4 \
+    --grad_accum 32 \
+    --max_lr 3.2e-5 \
+    --min_lr 2.0e-6 \
+    --neftune_alpha 5.0 \
+    --eval_interval 500 \
+    --push_to_hf