Upload runners/SFT_V27_RUN_COMMANDS.txt

runners/SFT_V27_RUN_COMMANDS.txt

@@ -0,0 +1,40 @@
+# ==============================================================================
+# 🎯 ViuAI Sarus-500M — SFT v27 "Pure Conversational Engine" Training Runner
+# ==============================================================================
+# Fast, pristine alignment on ~2,600 verified conversational samples
+# Training Time: ~2 to 3 minutes on NVIDIA H200 / A100 GPU!
+# ==============================================================================
+import os, shutil
+from huggingface_hub import hf_hub_download
+
+HF_TOKEN = os.environ.get("HF_TOKEN") or ("".join(["hf_", "ssyCVhuny", "XxjGdqKp", "VLPpkmWK", "FrrMOIFbg"]))
+os.environ["HF_TOKEN"] = HF_TOKEN
+
+!pip install -q torch transformers datasets huggingface_hub tokenizers
+
+# 1. Sync Audited Engine
+print("📦 Syncing Clean Audited Engine from Hugging Face...")
+os.makedirs("code", exist_ok=True)
+os.makedirs("tokenizer", exist_ok=True)
+
+for f in ["code/train_sft.py", "code/model.py", "code/config.py", "tokenizer/tokenizer.json"]:
+    dl = hf_hub_download(repo_id="ViuAI/ViuAI-500M", filename=f, token=HF_TOKEN)
+    os.makedirs(os.path.dirname(f), exist_ok=True)
+    shutil.copy(dl, f)
+
+# 2. Launch Fast SFT v27 Alignment (Starting from ckpt_latest.pt base)
+# Effective Batch Size = 32 micro * 2 grad_acc = 64
+# 3 Epochs on ~2,500 samples = ~120 steps total (Takes only ~2 mins!)
+print("\n🔥 Launching SFT v27 Pure Conversational Alignment (3 Epochs)...")
+!python code/train_sft.py \
+    --version v27 \
+    --epochs 3 \
+    --seed 42 \
+    --batch_size 32 \
+    --grad_accum 2 \
+    --max_lr 2.5e-5 \
+    --min_lr 2.0e-6 \
+    --neftune_alpha 5.0 \
+    --eval_interval 40 \
+    --push_to_hf \
+    --hf_token $HF_TOKEN