mirror of
https://github.com/8b-is/unified-8b-is.git
synced 2026-11-22 00:30:00 +01:00
Upstream pull mirror of https://github.com/8b-is/unified-8b-is.git
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| AGENTS.md | ||
| alexiai-external | ||
| aurora-8b-pq-external | ||
| aurora-external | ||
| Cargo.toml | ||
| CROSS_REFERENCE.md | ||
| memnet-external | ||
| mlx-quant-external | ||
| README.md | ||
unified-8b-is — Sovereign AI Constellation
One repo, one math, zero compromise.
Projects Cross-Linked
| Project | Path | Description |
|---|---|---|
| alexiai | alexiai-external/ |
Rust core + GAIA-MLX-QUANT substrate + BitNet b1.58 ternary |
| aurora | aurora-external/ |
Balanced ternary + MEM |
| MLX-QUANT | mlx-quant-external/ |
Bare-metal Apple Silicon tensor quantization (18 architectural layers) |
| memnet | memnet-external/ |
Semantic routing + Ayevn wave tokens + SoulSignatures |
| aurora-8b-pq | aurora-8b-pq-external/ |
Ternary audio codec (4.2 kbps) + PQM training pipeline |
Shared Invariants
p/q = 2^20 = 1_048_576 # Universal denominator
ternary = {-1, 0, +1} # Universal spin space
loopback = localhost/127.0.0.1 # Universal sovereignty
Quick Start
# Build all Rust lanes
cargo build --workspace --release
# Run sovereignty check
cargo run --release -p alexiai -- doctor
# Prepare training data (includes HF datasets from train.8b.is)
python -m aurora_8b_pq.data.prepare --hf-datasets quantal-train-v2 music-quant-signs
# Train ternary audio codec
python -m aurora_8b_pq.train --config configs/default.yaml
# Quantize for ESP32-S3 deployment
python -m aurora_8b_pq.quantize --checkpoint checkpoints/best.pt
python -m aurora_8b_pq.export --checkpoint checkpoints/best.pt --out deploy/include/
Train via train.8b.is (Remote GPU)
The train.8b.is portal provides:
- RTX 6000 Ada 48 GB + Xeon Gold 5412U + 756 GB DDR5 ECC
- PQM (p/q + Magnitude) quantization with OEIS A161330 snowflake bounds
- HF dataset mirrors:
PeetPedro/quantal-train-v2,PeetPedro/music-quant-signs,PeetPedro/quantal-train-assets - Golden-Logits gate verification + attestal.proof.v1 seals
To train remotely:
# 1. Upload your data to train.8b.is (requires Operator Key)
# 2. Launch PQM training via web UI or API
# 3. Download verified checkpoint + manifest (SHA-256 sealed)
# 4. Deploy via export.py
Cross-Link Map
alexiai.kernel.pq ──▶ aurora-core.Rational ──▶ MLX-QUANT.pq_scale
alexiai.kernel.ternary ──▶ aurora-core.TernarySpin ──▶ aurora-8b-pq.ternary_weights
alexiai.sovereign.guard ──▶ aurora.loopback_check ──▶ memnet.SoulSig.verify
alexiai.wupz.pack ──▶ aurora.mem8.capsule ──▶ memnet.PacketBatcher
alexiai.gaia.fold ──▶ aurora.phoenix.dyad ──▶ aurora-8b-pq.codebook
License
MIT OR Apache-2.0 across all projects.
the constellation · 0 + 1 · fine touch from within · vaked.dev