Reliable post-training for open sparse MoE models.
Carve open Mixture-of-Experts checkpoints into visual expert sectors, forensically inspect what each sector contains (with honest evidence tiers), train with specialist / broad / wide postures, enforce per-sector data contracts, promote gated AetherPackages — and never confuse dry-run CI completeness with weight-level MoE readiness.
v0.5.1 — PolyForm Noncommercial · complete operator guide · evidence tiers · sequential sector workflow · fused-expert PEFT · Training Console
© 2026 AetherAwareness — free for noncommercial use; not free to monetize (COMMERCIAL.md).
| Code | github.com/AetherAwareness/aetherforge |
| Download | Releases (source zip/tarball) |
| Complete guide | docs/GUIDE.md — setup, run, utilize, all postures |
| How-to cookbook | docs/HOWTO.md — task recipes |
| Product deep-dive | docs/product.md |
| User docs site | AetherAwareness.github.io/aetherforge |
Sparse MoE models only fire a thin slice of parameters per token. Generic fine-tuning tools:
- treat them like dense models, or
- miss fused expert banks used by many modern MoEs, or
- give no map of which capacity you are editing.
| Capability | Benefit |
|---|---|
| Expert Group Studio | Carve the lattice into ~active-fire sectors |
| Sector forensics | Mass, depth role, content signature — with evidence tiers |
| Sequential sector workflow | Forensics → dataset → ESFT per sector (siblings frozen) |
| Data contracts | Min real mass, max synth fraction, uniqueness gates |
| Plan freeze | Immutable membership fingerprint for each train wave |
| Probe keep/rollback | Pre/post routing share; rollback on regression |
| Postures | specialist · broad · wide lattice LoRA |
| Fused-expert PEFT | target_parameters + expert-index grad masks where needed |
| Scorecard | CI completeness vs MoE reliability (honest labels) |
| Remote train | Vast / RunPod / SSH — plan → sync → nohup → pull |
| Domain packs | Any industry — no hard-coded field bleed |
Full dense-style retrain of entire multi-hundred-billion MoE lattices is not the product goal. AetherForge trains adapters (ESFT/LoRA) on open checkpoints (optional selective full-expert updates).
Model coverage: AetherForge targets open sparse MoEs in general (module-list experts and fused expert banks). Built-in family profiles and recipes cover common shapes; you can add capacity/config for other MoEs via YAML.
git clone https://github.com/AetherAwareness/aetherforge.git
cd aetherforge
bash scripts/install.sh
source .venv/bin/activate
aetherforge quickstart # doctor + smoke dry-run + next steps
aetherforge status
aetherforge recipes
aetherforge dashboard # http://127.0.0.1:8765/aetherforge train --recipe dryrun --dry-run
aetherforge train --recipe broad-flash --dry-run
aetherforge init logistics --posture broad
aetherforge train --recipe broad-flash -c configs/domains/logistics.yaml --dry-run
aetherforge validate-flash # prove fused-expert PEFT stack
aetherforge workflow -c configs/base.yaml -c recipes/generic_dryrun.yaml --plan-only --dry-runWhen training.sector_mode: sequential (default), each train-enabled sector is its own mini-pipeline:
- Freeze plan fingerprint — membership cannot silently change mid-wave
- Forensic assess — evidence tier + calibrated themes
- Readiness gate — warn / block / skip
- Sector dataset + data contract
- Pre-probe routing mass
- ESFT only that sector’s experts; siblings frozen
- Post-probe → keep or rollback
- Interference summary across sectors
aetherforge train -c configs/base.yaml -c recipes/generic_dryrun.yaml --dry-run
aetherforge train --recipe dryrun --sector-mode joint --dry-run # single-pass ESFT| Tier | Meaning |
|---|---|
| structure_only | Geometry / mass / depth — no content claim |
| assignment | Operator or pack bound domain/topics/keywords |
| routing_probed | Affinity/routing matrix present (may be synthetic in dry-run) |
| Posture | Coverage |
|---|---|
| specialist | Few experts / selected sectors, one domain |
| broad | Many experts + multi-sector + multi-corpus |
| wide | Lattice-scale LoRA (still PEFT) |
Details: docs/guides/postures.md
| Label | Meaning |
|---|---|
| ci_completeness | Pipeline, data, proxies passed (includes dry-run) |
| moe_reliability | Live model + non-synthetic routing (+ sector keep when wave ran) |
Dry-run packages are stamped DRY_RUN_NOT_MOE_READY. Do not treat dry-run promotion as weight-level specialization.
aetherforge groups --preview --family generic_moe --num-groups 12
aetherforge forensics --family generic_moe --num-groups 12 --markdown
aetherforge dashboard # themes: NEXUS · MATRIX · PLASMA · AURORAIndustry content lives only in domain packs (never hard-coded in the trainer).
Many modern MoEs store routed experts as fused parameter banks rather than a simple ModuleList of MLPs. AetherForge applies PEFT target_parameters plus expert-index grad masks so routed capacity is trainable—not only shared experts.
aetherforge validate-flash
aetherforge train -c configs/base.yaml -c configs/<moe_family_profile>.yaml \
-c recipes/flagship_flash_domain.yaml --dry-runOptional YAML family profiles under configs/ illustrate common shapes; the product itself is any open MoE you can describe in capacity + PEFT config. See docs/guides/flash-0731.md for the fused-expert PEFT pattern.
Hardware: multi-GPU PEFT is the realistic path for large MoEs. Full bf16 Adam over every parameter of a frontier-scale sparse model is not the product goal.
| Stage | Output |
|---|---|
| diagnostics | model summary |
| data | train/eval + fingerprint |
| affinity | routing probe + selection |
| groups | sectors + forensics + readiness |
| esft | sequential sector wave or joint adapter |
| router_hygiene | router calibration |
| preference | preference / THD pairs |
| lifecycle | elastic expert plan |
| scorecard | CI vs MoE labels |
| package | AetherPackage (+ promote stamps) |
aetherforge doctor | validate | validate-flash | train | data | workflow | probe
aetherforge groups | forensics | dashboard | runs | status | recipes | init | quickstart
aetherforge connect | remote
aetherforge scorecard | package | consult
Keys stay in ~/.aetherforge/ (not the git tree).
export VAST_API_KEY=…
aetherforge connect key vast --from-env
aetherforge connect vast --host HOST --port PORT
aetherforge remote launch --exec --recipe broad-flash
aetherforge remote pull && aetherforge remote logs --tail 80| Doc | Description |
|---|---|
| Complete guide | What / why / setup / run / utilize / hardware |
| Product explanation | Thorough product deep-dive |
| GitHub Pages | User guide site |
| Getting started | Install & first dry-run |
| Architecture | Pipeline & MoE concepts |
| Studio & forensics | Lattice + sector dossiers |
| Fused-expert PEFT | Fused banks & PEFT pattern |
| Postures | specialist / broad / wide |
| Safety | High-stakes & privacy |
| Commercial use | What is / is not monetizable |
| Changelog | Versions |
eval:
scorecard_thresholds:
high_stakes: true
require_human_approval: true
domain_depth_min: 0.65pip install -e ".[dev]"
pytest tests/ -q- Dry-run ≠ trained MoE. Synthetic affinity and CI scorecards prove the factory, not weight-level specialization.
- Live multi-GPU PEFT is the real train path for large sparse models.
- Domain packs supply industry content — the core never hard-codes a field.
- Secrets belong in
~/.aetherforge/, never in the repo.
Copyright © 2026 AetherAwareness.
AetherForge is licensed under the PolyForm Noncommercial License 1.0.0.
| You may | You may not (without a commercial grant) |
|---|---|
| Use, study, modify for noncommercial purposes | Sell AetherForge or a paid fork |
| Share with license + Required Notice | Offer paid SaaS / managed train service built on it |
| Research, education, hobby, public knowledge | Bundle into a commercial product for sale |
See NOTICE and COMMERCIAL.md.
Trademark/names “AetherForge”, “AetherPackage”, “AetherAwareness” remain with AetherAwareness.
Contact: admin@aetherawareness.com