Create todo.md

llm/todo.md

@@ -0,0 +1,46 @@
+ Step 1: Install dependencies
+
+     pip install torch transformers safetensors tqdm
+
+     Step 2: Test circuit execution standalone
+
+     cd C:\Users\cnort\8bit-threshold-computer\llm
+     python circuit_llm.py --circuit-path ../neural_computer.safetensors --device cpu
+
+     Expected output:
+     - Load SmolLM2-360M
+     - Baseline arithmetic evaluation (~1-10% accuracy)
+     - Augment layers 10-20 with threshold circuits
+     - Test circuit execution (127+128=255, etc.)
+
+     Step 3: Train interface layers (CPU)
+
+     python train_circuit_interface.py \
+       --circuit-path ../neural_computer.safetensors \
+       --device cpu \
+       --epochs 3 \
+       --batch-size 4 \
+       --n-samples 2000
+
+     Expected:
+     - Freeze 362M params
+     - Train ~1.37M interface params (BitExtractor, BitInjector, Router)
+     - 3 epochs on arithmetic examples (reduced samples for CPU)
+     - Accuracy should improve from ~10% to >90%
+     - Training time: ~1-2 hours on CPU
+
+     Step 4: Save trained model
+
+     Output: circuit_augmented_smollm2.pt
+
+     Hardware Requirements
+
+     - GPU recommended (RTX 6000 Ada per guide.md)
+     - CPU possible but slow (~10x longer)
+     - ~4GB VRAM for SmolLM2-360M
+
+     Success Criteria
+
+     - Baseline accuracy: ~1-10% (SmolLM2 guessing)
+     - Post-training accuracy: >98% (circuits computing)
+     - Circuit tests pass: 127+128=255, 255+1=0, etc.