Skip to content

TUTORIAL.md

Latest commit

 

History

History
47 lines (40 loc) · 2.48 KB

TUTORIAL.md

File metadata and controls

47 lines (40 loc) · 2.48 KB

Mamba-3 Latent Engine Tutorial

Welcome to the cutting-edge continuous state reasoning pipeline. This tutorial covers how to deploy the Phase 11 through Phase 13 curriculum to force a sub-1B parameter Mamba model to reason systematically in the dark, bypassing explict Chain-of-Thought token generation.

1. Prerequisites & Data Synthesis

Before jumping into the Latent Forge, you must construct the specialized curriculums:

python phase12_curriculum_builder.py

This script generates the required spacing-hacked arithmetic sets to protect mathematical geometry from BPE string tokenization destruction.

  • phase12a_alu.jsonl: 15,000 pure arithmetic combinations.
  • phase12b_gsm8k.jsonl: 7,473 curated word problems stripped of CoT formatting.

2. Phase 12-A: The ALU Burn-in

Train the basic numeric syntax and string addition behaviors into the Mamba matrix parameters.

python phase12a_sft_trainer.py

Key Feature: The TURING OVERRIDE abort sequence. The model will automatically freeze and save the weights (mamba3_p12A_alu.pt) if the moving average training loss crosses < 0.15 or if validation stalls. This prevents catastrophic overfitting to arithmetic arrays.

3. Phase 12-B: The Semantic Bridge

Link the abstract mathematics to the linguistic English parsing.

python phase12b_sft_bridge.py

Key Feature: Uses a mixed 80/10/10 Replay Buffer natively interleaving Phase 11 boolean logic gates, Phase 12-A ALU calculations, and Target-Masked English Word queries (labels=-100 over the prose). Aborts automatically after the action space collapses.

4. Phase 12-C: Final Boss GRPO

Forge the mathematical continuous constraints using Group Relative Policy Optimization.

python mamba3_p12_grpo.py

This script utilizes an $N=8$ branch span and unrolls dynamic continuous loops. Because the Phase 12-B format floor is solid, the specific sub-branches that calculate perfect arithmetic in the dark will receive massive positive advantage distributions.

5. Phase 13: Conversational Re-Anchoring

Recover English generation.

python phase13_conversational_reanchoring.py

Because the engine is permanently target-masked during Phase 12, Phase 13 re-injects standard SFT chat sequences at a 50/50 ratio to build a structurally unified model capable of both math mapping and casual conversation.

Running Tests

To verify the latent routing logic against the conversational english logic, run:

python test_general.py
python test_turing.py