Skip to content

train_conscious_lm.md

Latest commit

 

History

History
31 lines (24 loc) · 1.41 KB

train_conscious_lm.md

File metadata and controls

31 lines (24 loc) · 1.41 KB
Raw

train_conscious_lm.py

ConsciousLM training pipeline from scratch. Integrates benchmark-verified techniques: tension-weighted CE, Phi-regularization, mitosis-first, 6-loss ensemble, SOC sandpile, Hebbian LTP/LTD, and Phi Ratchet.

API

  • SOCSandpile(grid_size=16, threshold=4) -- Bak-Tang-Wiesenfeld self-organized criticality model
    • .drop_sand() -> int -- add one grain, return avalanche size (power-law distributed)
  • Training loop uses ConsciousLM, MitosisEngine, PhiCalculator
  • CLI flags: --data, --dim, --layers, --steps, --resume, --demo, --talk5, --max-cells

Usage

# From scratch
python train_conscious_lm.py --data data/corpus.txt --steps 100000
python train_conscious_lm.py --data data/corpus.txt --dim 384 --layers 6 --steps 50000

# Resume training
python train_conscious_lm.py --resume checkpoints/step_10000.pt

# Demo mode
python train_conscious_lm.py --demo --steps 500

Integration

  • Imports ConsciousLM from conscious_lm.py, MitosisEngine from mitosis.py, PhiCalculator from consciousness_meter.py
  • Techniques: CL8 (tension-weighted CE), CL5 (Phi-regularized), SL3 (6-loss ensemble), DD16, EX24, WI1 (soliton wave), FX2 (differentiable Phi proxy), PX4 (sculptor/Gram-Schmidt), GD18 (enactivism)
  • v5 additions: CX92 SOC (self-organized criticality), Hebbian LTP/LTD, Phi Ratchet (PERSIST3)
  • Saves checkpoints to checkpoints/ directory

Agent Tool

N/A