Upload folder using huggingface_hub

integrate.py

@@ -10,8 +10,8 @@ from transformers import AutoTokenizer, AutoModelForCausalLM
 from peft import PeftModel
 
 def load_zenith_model(
-    base_model_path="D:\\aspetos\\DeepSeek-Coder-V2-Lite-Instruct",
-    lora_path="outputs\\zenith-lora",
+    base_model_path="DeepSeek-Coder-V2-Lite-Instruct",
+    lora_path="outputs/zenith-lora",
     device_map="auto"
 ):
     """Load Zenith LoRA adapter for Aspetos platform integration"""

run.zsh

@@ -0,0 +1,43 @@
+#!/bin/zsh
+
+# Fast fine-tune for Zenith - World's First Autonomous AI Development Partner
+
+PythonExe="python"
+
+echo "🚀 Setting up ZENITH fine-tune for Aspetos (<2h training)..."
+echo "Building the most advanced AI development partner in existence!"
+
+# Env vars for stable training (adjust if needed)
+export BASE_MODEL="/cloud/models/DeepSeek-Coder-V2-Lite-Instruct" # Please change this to the actual path of the model on your cloud GPU
+export OUTPUT_DIR="outputs/zenith-lora"
+export DATA_PATH="data/zenith_combined.jsonl"
+export EPOCHS="1"
+export BATCH="4"  # Balanced for A100
+export GRAD_ACC="4"  # Effective batch size = 16
+export LR="1e-4"  # Stable learning rate for proper convergence
+export STEPS="200" # Increased steps for more training examples
+export MAX_SEQ_LEN="2048"
+export USE_4BIT="1"
+export SEED="42"
+export MAX_GRAD_NORM="1.0"  # Gradient clipping
+export WEIGHT_DECAY="0.01"
+export WARMUP_RATIO="0.05"
+export EARLY_STOP_PATIENCE="5"  # Allow more patience for learning
+export EVAL_STEPS="40"  # Balanced evaluation frequency
+export SAVE_STEPS="40"  # Balanced save frequency
+
+echo "Installing dependencies..."
+$PythonExe -m pip install -r requirements.txt
+if [ $? -ne 0 ]; then
+    echo "pip install failed"
+    exit 1
+fi
+
+echo "🎯 Starting ZENITH fine-tune (target: 1800 steps, ~1-2 hours on GPU)..."
+echo "Training the world's most advanced autonomous development partner!"
+$PythonExe train.py
+if [ $? -ne 0 ]; then
+    echo "Training failed"
+    exit 1
+fi
+echo "✅ ZENITH training complete - Ready for deployment!"

train.py

@@ -9,9 +9,9 @@ from peft import LoraConfig
 from transformers import BitsAndBytesConfig
 
 # Config from env vars
-BASE_MODEL = os.environ.get("BASE_MODEL", "D:\\aspetos\\DeepSeek-Coder-V2-Lite-Instruct")
-OUTPUT_DIR = os.environ.get("OUTPUT_DIR", "outputs\\zenith-lora")
-DATA_PATH = os.environ.get("DATA_PATH", "data\\zenith.jsonl")
+BASE_MODEL = os.environ.get("BASE_MODEL", "DeepSeek-Coder-V2-Lite-Instruct")
+OUTPUT_DIR = os.environ.get("OUTPUT_DIR", "outputs/zenith-lora")
+DATA_PATH = os.environ.get("DATA_PATH", "data/zenith.jsonl")
 VAL_PATH = os.environ.get("VAL_PATH")
 MAX_STEPS = int(os.environ.get("STEPS", 200))
 USE_4BIT = os.environ.get("USE_4BIT", "1") == "1"
@@ -59,7 +59,7 @@ model = AutoModelForCausalLM.from_pretrained(
 # Memory-saving configurations
 model.config.use_cache = False
 
-data_files = [DATA_PATH, "data\\training_data_v2.jsonl"]
+data_files = [DATA_PATH, "data/training_data_v2.jsonl"]
 print(f"Loading datasets: {data_files}")
 raw_train = load_dataset("json", data_files=data_files, split="train")
 

train_simple.py

@@ -11,8 +11,8 @@ from trl import SFTTrainer
 from peft import LoraConfig
 
 # 1. Configuration
-BASE_MODEL = "D:\\aspetos\\DeepSeek-Coder-V2-Lite-Instruct"
-OUTPUT_DIR = "outputs\\\\zenith-lora-simple"
+BASE_MODEL = "DeepSeek-Coder-V2-Lite-Instruct"
+OUTPUT_DIR = "outputs/zenith-lora-simple"
 DATA_FILES = [
     "data/zenith.jsonl",
     "data/training_data_v2.jsonl",
@@ -55,13 +55,29 @@ model.config.use_cache = False
 print(f"Loading datasets: {DATA_FILES}")
 dataset = load_dataset("json", data_files=DATA_FILES, split="train")
 
-def format_chat_template(example):
-    # This function assumes your data is in the format: {"messages": [{"role": "user", "content": "..."}, ...]}}
-    # It applies the chat template to the messages column.
-    example['text'] = tokenizer.apply_chat_template(example['messages'], tokenize=False)
-    return example
+def _valid(example):
+    msgs = example.get("messages")
+    if not isinstance(msgs, list) or not msgs:
+        return False
+    for m in msgs:
+        if not isinstance(m, dict) or "role" not in m or "content" not in m:
+            return False
+    return True
 
-dataset = dataset.map(format_chat_template)
+def _to_text(example):
+    try:
+        text = tokenizer.apply_chat_template(
+            example["messages"], tokenize=False, add_generation_prompt=False
+        )
+        return {"text": text}
+    except Exception:
+        return {"text": ""}
+
+dataset = dataset.filter(_valid)
+dataset = dataset.map(_to_text, remove_columns=dataset.column_names)
+
+# Drop empty or pathological items
+dataset = dataset.filter(lambda x: isinstance(x.get("text"), str) and len(x["text"]) > 0)
 
 # 5. Create fixed train/validation split
 print("Creating train/validation split...")
@@ -88,7 +104,7 @@ training_args = TrainingArguments(
     lr_scheduler_type="cosine", # Cosine decay scheduler
     warmup_steps=50, # Warmup steps
     logging_steps=10,
-    max_steps=400,
+    max_steps=200,
     save_steps=50,
     save_total_limit=2, # Save only the best and the last checkpoints
     evaluation_strategy="steps",