Skip to content

HOWTO.md

Latest commit

 

History

History
126 lines (91 loc) · 2.96 KB

HOWTO.md

File metadata and controls

126 lines (91 loc) · 2.96 KB
Raw
title How-to cookbook
layout default
nav_order 3

AetherForge — How-to Cookbook

{: .no_toc }

Task-oriented recipes. For the full story see GUIDE.md and product.md. {: .fs-6 .fw-300 }

Table of contents

{: .no_toc .text-delta }

  1. TOC {:toc}

Install from release

curl -sL https://github.com/AetherAwareness/aetherforge/archive/refs/tags/v0.5.1.tar.gz | tar xz
cd aetherforge-0.5.1
bash scripts/install.sh
source .venv/bin/activate
aetherforge doctor

First dry-run (prove the factory)

aetherforge train --recipe dryrun --dry-run
aetherforge dashboard   # http://127.0.0.1:8765/

Expect under artifacts/runs/…: sector_forensics.json, plan_freeze.json, sector_workflow/, PROMOTION_LABEL.txt.

Scaffold a domain pack

aetherforge init my_field --posture broad
# edit configs/domains/my_field.yaml
aetherforge train -c configs/base.yaml -c configs/domains/my_field.yaml --dry-run

Sector forensics only

aetherforge groups --preview --family generic_moe --num-groups 12
aetherforge forensics --family generic_moe --num-groups 12 --markdown
aetherforge workflow -c configs/base.yaml -c recipes/generic_dryrun.yaml --plan-only --dry-run

Partition data into sector shards

aetherforge data -c configs/base.yaml -c recipes/generic_dryrun.yaml --sectors --dry-run

Specialist vs broad vs wide

aetherforge train --recipe dryrun --dry-run
aetherforge train --recipe broad-flash --dry-run
aetherforge train --recipe wide-flash --dry-run

Fused-expert PEFT stack check

aetherforge validate-flash
aetherforge train -c configs/base.yaml -c configs/<moe_family_profile>.yaml \
  -c recipes/flagship_flash_domain.yaml --dry-run

Remote GPU train (outline)

aetherforge connect key vast --from-env   # after export VAST_API_KEY
aetherforge connect vast --host HOST --port PORT
aetherforge remote plan --recipe broad-flash
aetherforge remote launch --exec --recipe broad-flash
aetherforge remote logs --tail 100
aetherforge remote pull

High-stakes human promote

# in domain or run config
eval:
  scorecard_thresholds:
    high_stakes: true
    require_human_approval: true

Then open aetherforge dashboard → Approve / Reject / Force promote.

Read run artifacts

File Meaning
sector_forensics.md What each sector contains + evidence tier
plan_freeze.json Immutable membership hash for the wave
sector_workflow/ Per-sector datasets, probes, keep/rollback
scorecard.json CI vs MoE labels
promoted/DRY_RUN_NOT_MOE_READY.txt Dry-run honesty stamp

After training: serve the model

  1. Convert/quantize adapter to your serving format (e.g. GGUF) as you prefer
  2. Serve with your usual local stack (llama-server, Ollama, vLLM, etc.)

Tests

pytest tests/ -q

Contact

admin@aetherawareness.com · PolyForm Noncommercial · © 2026 AetherAwareness