Upload training_scripts/quickstart.py with huggingface_hub
Browse files- training_scripts/quickstart.py +184 -0
training_scripts/quickstart.py
ADDED
|
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
ÜBERMENSCHETIEN QUICK START
|
| 4 |
+
============================
|
| 5 |
+
One-command setup and training.
|
| 6 |
+
|
| 7 |
+
Usage:
|
| 8 |
+
python quickstart.py --full # Run full pipeline
|
| 9 |
+
python quickstart.py --train-dense # Just dense training
|
| 10 |
+
python quickstart.py --train-cfhot # Just CF-HoT heads
|
| 11 |
+
python quickstart.py --improve # Just self-improvement
|
| 12 |
+
python quickstart.py --test # Test current model
|
| 13 |
+
|
| 14 |
+
"From zero to self-improving in one command"
|
| 15 |
+
"""
|
| 16 |
+
|
| 17 |
+
import os
|
| 18 |
+
import sys
|
| 19 |
+
import argparse
|
| 20 |
+
import subprocess
|
| 21 |
+
from pathlib import Path
|
| 22 |
+
|
| 23 |
+
ROOT = os.path.dirname(os.path.abspath(__file__))
|
| 24 |
+
|
| 25 |
+
def run_command(cmd, description):
|
| 26 |
+
"""Run a command with nice output."""
|
| 27 |
+
print(f"\n{'='*70}")
|
| 28 |
+
print(f"🚀 {description}")
|
| 29 |
+
print(f"{'='*70}")
|
| 30 |
+
print(f"$ {cmd}\n")
|
| 31 |
+
|
| 32 |
+
result = subprocess.run(cmd, shell=True, cwd=ROOT)
|
| 33 |
+
|
| 34 |
+
if result.returncode != 0:
|
| 35 |
+
print(f"\n❌ Failed: {description}")
|
| 36 |
+
return False
|
| 37 |
+
|
| 38 |
+
print(f"\n✓ Complete: {description}")
|
| 39 |
+
return True
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def check_dependencies():
|
| 43 |
+
"""Check required packages are installed."""
|
| 44 |
+
print("\n🔍 Checking dependencies...")
|
| 45 |
+
|
| 46 |
+
required = [
|
| 47 |
+
"torch",
|
| 48 |
+
"transformers",
|
| 49 |
+
"peft",
|
| 50 |
+
"bitsandbytes",
|
| 51 |
+
"accelerate",
|
| 52 |
+
]
|
| 53 |
+
|
| 54 |
+
missing = []
|
| 55 |
+
for pkg in required:
|
| 56 |
+
try:
|
| 57 |
+
__import__(pkg)
|
| 58 |
+
print(f" ✓ {pkg}")
|
| 59 |
+
except ImportError:
|
| 60 |
+
print(f" ✗ {pkg}")
|
| 61 |
+
missing.append(pkg)
|
| 62 |
+
|
| 63 |
+
if missing:
|
| 64 |
+
print(f"\n❌ Missing packages: {', '.join(missing)}")
|
| 65 |
+
print("Install with: pip install " + " ".join(missing))
|
| 66 |
+
return False
|
| 67 |
+
|
| 68 |
+
print("\n✓ All dependencies installed")
|
| 69 |
+
return True
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def train_dense(steps=100):
|
| 73 |
+
"""Run THE CONDENSATOR dense training."""
|
| 74 |
+
return run_command(
|
| 75 |
+
f"python the_condensator.py --stages sft,dpo,rl --steps {steps}",
|
| 76 |
+
"THE CONDENSATOR - Dense Response Training"
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def train_cfhot(steps=3000):
|
| 81 |
+
"""Train CF-HoT behavior heads."""
|
| 82 |
+
success = True
|
| 83 |
+
|
| 84 |
+
for behavior in ["repetition", "hedging", "verbosity"]:
|
| 85 |
+
if not run_command(
|
| 86 |
+
f"python train_cfhot_head.py --behavior {behavior} --steps {steps}",
|
| 87 |
+
f"CF-HoT {behavior.upper()} Head Training"
|
| 88 |
+
):
|
| 89 |
+
success = False
|
| 90 |
+
|
| 91 |
+
return success
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
def train_self_improve(iterations=5):
|
| 95 |
+
"""Run stable self-improvement."""
|
| 96 |
+
return run_command(
|
| 97 |
+
f"python train_self_improve.py --iterations {iterations}",
|
| 98 |
+
"Stable Self-Improvement Loop"
|
| 99 |
+
)
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def test_model(checkpoint=None):
|
| 103 |
+
"""Test the model."""
|
| 104 |
+
cmd = "python the_condensator.py --eval-only"
|
| 105 |
+
if checkpoint:
|
| 106 |
+
cmd += f" --checkpoint {checkpoint}"
|
| 107 |
+
return run_command(cmd, "Model Evaluation")
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def full_pipeline():
|
| 111 |
+
"""Run the complete training pipeline."""
|
| 112 |
+
print("\n" + "="*70)
|
| 113 |
+
print("🔥 ÜBERMENSCHETIEN FULL TRAINING PIPELINE")
|
| 114 |
+
print("="*70)
|
| 115 |
+
print("""
|
| 116 |
+
This will run:
|
| 117 |
+
1. THE CONDENSATOR (SFT → DPO → RL)
|
| 118 |
+
2. CF-HoT Head Training (repetition, hedging, verbosity)
|
| 119 |
+
3. Stable Self-Improvement Loop
|
| 120 |
+
|
| 121 |
+
Estimated time: 2-4 hours on RTX 3090
|
| 122 |
+
""")
|
| 123 |
+
|
| 124 |
+
if not check_dependencies():
|
| 125 |
+
return False
|
| 126 |
+
|
| 127 |
+
# Step 1: Dense training
|
| 128 |
+
if not train_dense(100):
|
| 129 |
+
return False
|
| 130 |
+
|
| 131 |
+
# Step 2: CF-HoT heads
|
| 132 |
+
if not train_cfhot(1000): # Fewer steps for quick start
|
| 133 |
+
return False
|
| 134 |
+
|
| 135 |
+
# Step 3: Self-improvement
|
| 136 |
+
if not train_self_improve(3):
|
| 137 |
+
return False
|
| 138 |
+
|
| 139 |
+
print("\n" + "="*70)
|
| 140 |
+
print("✓ ÜBERMENSCHETIEN TRAINING COMPLETE!")
|
| 141 |
+
print("="*70)
|
| 142 |
+
print("""
|
| 143 |
+
Your model is ready! Run:
|
| 144 |
+
|
| 145 |
+
python ubermenschetien_v2_full.py
|
| 146 |
+
|
| 147 |
+
Commands:
|
| 148 |
+
> hello # Chat
|
| 149 |
+
> !eval # Evaluate quality
|
| 150 |
+
> !improve # Continue self-improvement
|
| 151 |
+
""")
|
| 152 |
+
|
| 153 |
+
return True
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
def main():
|
| 157 |
+
parser = argparse.ArgumentParser(description="Übermenschetien Quick Start")
|
| 158 |
+
parser.add_argument("--full", action="store_true", help="Run full pipeline")
|
| 159 |
+
parser.add_argument("--train-dense", action="store_true", help="Run dense training only")
|
| 160 |
+
parser.add_argument("--train-cfhot", action="store_true", help="Run CF-HoT training only")
|
| 161 |
+
parser.add_argument("--improve", action="store_true", help="Run self-improvement only")
|
| 162 |
+
parser.add_argument("--test", action="store_true", help="Test current model")
|
| 163 |
+
parser.add_argument("--steps", type=int, default=100, help="Training steps")
|
| 164 |
+
parser.add_argument("--checkpoint", type=str, default=None, help="Checkpoint path for testing")
|
| 165 |
+
|
| 166 |
+
args = parser.parse_args()
|
| 167 |
+
|
| 168 |
+
if args.full:
|
| 169 |
+
full_pipeline()
|
| 170 |
+
elif args.train_dense:
|
| 171 |
+
train_dense(args.steps)
|
| 172 |
+
elif args.train_cfhot:
|
| 173 |
+
train_cfhot(args.steps)
|
| 174 |
+
elif args.improve:
|
| 175 |
+
train_self_improve()
|
| 176 |
+
elif args.test:
|
| 177 |
+
test_model(args.checkpoint)
|
| 178 |
+
else:
|
| 179 |
+
parser.print_help()
|
| 180 |
+
print("\n💡 Try: python quickstart.py --full")
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
if __name__ == "__main__":
|
| 184 |
+
main()
|