Ms.MoE Maker

Multi-Specified Mixture of Experts. Five deliberate experts instead of a hundred lottery tickets.

The design thesis is the inverse of a frontier MoE. Instead of training many experts and hoping specialisation emerges — then fighting dead and collapsed experts with a load-balancing auxiliary loss — you hand-assign the domains. Every expert has a guaranteed constituency, so none of them can go dead, because none of them was speculative.

The corollary is what makes it maintainable by one person: because each expert does exactly one thing, you can retrain one and re-splice without touching the others.

Not a coding model. A coding model shaped like your stack.

The real product is the factory, not the model. Swap the expert list and someone else gets their own Ms.MoE.

Install

pip install ms-moe-maker            # recipe-side: describe, validate, plan
pip install ms-moe-maker[train]     # on the box that actually builds

The base install depends only on pyyaml, and that is the point: ms-moe-maker validate and build --plan run on a laptop with no GPU, so you can check a recipe and read what it will cost before going near a machine that can run it. Every torch import in the package is inside a function, so the base install imports cleanly.

[train] adds torch, transformers, datasets, safetensors, accelerate, peft and trl — everything the build stages actually touch. The stitcher is vendored, so there is nothing else to check out alongside it.

Who needs which

The machine that trains does not have to be the machine you use. Rent a GPU for the expensive part, or have someone hand you the specialists, and the rest still works:

The second row is the one worth knowing about. export and smoke shell out to llama.cpp and import nothing but the standard library — so a person handed a finished MoE can convert it to GGUF and prove it generates, with no ML stack installed at all. Release CI asserts that, so it stays true.

Use

# Start from nothing
ms-moe-maker init > recipe.yaml
ms-moe-maker init --template dnd > recipe.yaml

# Discover what's available (zero side-effects, returns JSON)
ms-moe-maker describe

# Validate recipe structure — no pipeline, no GPU, no network
ms-moe-maker validate recipe.yaml

# Resolve config + stage plan and run NOTHING. Also no GPU.
ms-moe-maker build recipe.yaml --plan

# Run the full pipeline (needs torch, GPU, training venv)
ms-moe-maker build recipe.yaml

# JSON Lines events on stdout, prose on stderr
ms-moe-maker build recipe.yaml --json

# A real build on the smallest rung — cheap, but still a build
ms-moe-maker build recipe.yaml --dryrun

# Smoke-test the GGUF — checks it generates real tokens
ms-moe-maker smoke recipe.yaml

# Does the router prefer each expert on its own ground? (dead-expert check)
ms-moe-maker eval recipe.yaml --mode routing

# Does it answer better than one expert alone?
ms-moe-maker eval recipe.yaml --mode quality

--plan and --dryrun differ on purpose. --plan resolves everything and runs nothing, so it works on a laptop. --dryrun is a real build on the smallest rung — cheap, but it needs torch like any other build.

The recipe

A build, as a document. You can hand it to someone who doesn't have your box and they get your run.

Minimal recipe

schema_version: 1
name: my-moe

experts:
  - name: python
    source: { kind: stack, language: Python }
  - name: csharp
    source: { kind: stack, language: C# }

size: auto

That's it. The rest auto-fills from your hardware tier:

If you omit size, the tier's default is used. If you omit the tier, the middle tier (xavier) is the default.

A small run that is still a real one

The corpus volume defaults to production size. To watch the whole flow finish this evening — every stage, real artifacts, a loadable GGUF — turn the volume down rather than reaching for --dryrun:

corpus:
  min_samples: 300          # fail the stage below this, not train on scraps
  max_samples: 3000         # cap per expert
  router_mix_total: 800     # rows in the router's stratified mix

budget:
  target_steps: 150
  max_seq_length: 1024

One thing to know before it surprises you: min_samples is a floor, not a target, and it rises to meet router_mix_total — the router's mix is drawn from the training split, so each expert needs enough collected documents to fill its share of the mix. The run says so when it happens. See corpus: below.

--dryrun is a different thing: it also relabels the run and writes to a separate directory, because its job is structural testing, not a small build. ms-moe-maker build recipe.yaml --plan prints the volume back at you, so you can tell which of the two you are about to start.

See recipe.flow-0.5B.yaml for a complete worked example.

Defaults, and setting a box up for someone else

Every knob below has a default. Those defaults live in a file, not in the code, so you can configure a machine once — for yourself, or for someone you're handing it to — and every recipe on that box inherits it without saying a word.

A defaults file is just a recipe with no experts. Same keys, same blocks, same -1 sentinels, same typo warnings. Anything you can write in a recipe's budget: / corpus: / router: blocks, you can write here.

Start from a commented starter rather than from memory:

ms-moe-maker init --defaults-template          # writes ~/.msmoe/defaults.yaml
ms-moe-maker init --defaults-template --output -   # or just look at it

# ~/.msmoe/defaults.yaml — the box, set up once
budget:  { target_steps: 400 }
corpus:  { max_samples: 9000, router_mix_total: 8000 }
tools_expert: { teacher: Qwen/Qwen2.5-7B-Instruct }

# and then this is a whole recipe
schema_version: 1
name: sister-moe
size: 0.5B
tools_expert: true
experts:
  - name: python
    source: { kind: stack, language: Python }
  - name: markdown
    source: { kind: stack, language: Markdown }

Layers, later wins — and the recipe always wins:

-1 still means "you decide" at every layer, so it falls through instead of overwriting the layer below.

experts:, name: and template: are not accepted in a defaults file. Those describe one build, not a box.

Your hardware, and your checkpoints

Two blocks are the other way round — box only. A recipe may name a tier; it may not redefine one, or the same recipe would mean different hardware depending on who ran it.

tiers:
  spark:
    default_size: 14B        # this box would rather not reach for 32B
    default_lora_r: 96
  orin_agx:                  # a tier the tool has never heard of
    like: spark              # inherit the rest, change three things
    max_vram_gb: 64
    default_size: 7B
    default_quant: Q5_K_M

models:
  "0.5B": D:/models/Qwen2.5-Coder-0.5B-Instruct   # a local mirror
  "7B":
    safe: Qwen/Qwen2.5-Coder-7B
    abliterated: huihui-ai/Qwen2.5-Coder-7B-Instruct-abliterated

Adding hardware is the normal case, not the exotic one, so an unknown tier name is created, not refused. It needs either every field or a like: to inherit from — and if it's missing something, the run says which. Nothing here raises: a bad tier, an unknown field, a like: pointing at nothing, or a recipe asking for a tier this box doesn't have all come back as warnings.

--describe reports the tiers this install actually offers, so a ms-moe-maker build --plan tells you what your tier means here:

Ms.MoE — sister-moe  size=7B  tier=orin_agx
  tier     orin_agx: 64 GB, default 7B, lora_r 96, Q5_K_M
  base     /mnt/models/Qwen2.5-Coder-7B-Instruct-abliterated

Where did that number come from? validate and build --plan print every value that came from outside the recipe, and name the file:

DEFAULTS (5 from outside the recipe):
  budget.target_steps      <- ~/.msmoe/defaults.yaml
  corpus.max_samples       <- ~/.msmoe/defaults.yaml
  tools_expert.teacher     <- ~/.msmoe/defaults.yaml

Two ids, because there are two questions

Because defaults live on the box, a recipe on its own no longer fully determines a build. So there are two ids and they mean different things:

Both print in validate and build --plan. recipe_id is the recipe as written — it does not move when you change your box.

The run manifest records the build_id, the whole resolved configuration, and a hash of every defaults file that contributed. That is what lets a resumed build notice it is not the same build:

REFUSING TO RESUME: this run directory was built by a different build.
2 stage(s) already finished and would be kept as-is: data.corpus, finetune.python

What changed:
  · target_steps: 400 -> 1200
  · defaults file box.yaml: 4d38202c5581 -> 22de8b740060

Pick one:
  --force                 rebuild everything with the new settings
  --defaults <the old file>   reproduce the original build
  build somewhere else    change roots.output, keep both

Without that, changing a knob halfway through a build gives you a model whose specialists were trained differently from each other — silently, because finished stages skip themselves. It only refuses when something is already finished and would be inherited; a fresh directory just gets restamped.

Pass --defaults when you need two machines to agree, and read the provenance block when they don't.

ms-moe-maker --describe reports what a box presets — the layers, whether each is present, its hash, and which key came from which file — so a front-end can show a machine's configuration without re-implementing the merge. Under --json, validate and build emit a defaults event carrying the same provenance.

The full design and roadmap live in docs/DEFAULTS.md.

Every knob, and what actually moves

Everything below is optional. A recipe with nothing but experts: builds. This section is for when you want to turn something and would like to know what it does before you spend four hours finding out.

Where a number here is called measured, it came off a real run on a DGX Spark at 0.5B with three code experts, top-2 of 3 — not off a napkin. Where it isn't, it's a default someone picked, and you should feel free to disagree with it.

-1 means "you decide" on every numeric knob. That is not the same as 0. A recipe that omits a block behaves exactly like one that sets every field in it to -1, which is what lets us change a default without breaking your file.

budget: — how hard each specialist trains

Reach for target_steps and corpus.max_samples before you reach for lora_r. Measured: at 0.5B the rank was already 128 while each expert saw 1.23M tokens — one sixteenth of the rung that worked. A large adapter over a small corpus is what a 0.05-nat expert looks like.

Specialists are the cheap part to get right and the expensive part to run. The gate below is the opposite.

router: — how hard the gate trains

This is the block that decides whether your MoE routes at all. The stitch seeds the gate with small noise (router_init: random), so the router starts uniform-ish and has to learn everything it knows from this budget — and the budget's step count is the lever that actually moves enrichment.

The number that actually governs the gate is steps, and steps are not a knob — they're arithmetic:

router steps = router_mix_total × epochs ÷ (batch × accum)

Reps and sets: a mix row is a rep, an epoch is a set. epochs: 2 and router_mix_total: 8000 buy the same steps; the first is free, the second costs a corpus.

Measured dose-response — same stitch, same experts, only the router's step budget changed:

The fit across those four points is enrichment − 1 ≈ 0.0054 × steps^0.55, which is a rule of thumb worth carrying: doubling the router's steps buys about 46% more excess-over-1.0. It does not double anything. Input-dependence (JS divergence between routing distributions) moves the same way, about ×1.5 per doubling.

The ceiling for top-2 of 3 experts is 2.0× — if every token routed to its own expert plus one other, own-source share is 2/3 and the mean other share is 1/3. So 1.34× is roughly a third of the way to a gate that is perfectly opinionated, and no amount of steps takes you past 2.0 at this topology.

moe: — the shape of the stitched model

experts_per_tok: 1 is refused at validate, on purpose. With k=1 and norm_topk_prob: true, the single gate weight is divided by itself — it is always exactly 1.0, the language-model loss has no gradient path to the gate, and the router learns nothing while looking like it trained. With norm_topk_prob: false it's worse in a quieter way: the gate probability becomes a free scalar gain on a frozen expert, so the gate can lower the loss by adjusting volume rather than by choosing correctly. Neither is a router.

Setting experts_per_tok equal to your expert count is legal — it's a dense ensemble — but it makes the dead-expert measurement impossible rather than merely hard, because every expert is selected on every token by arithmetic.

router_init: zero is available for verifying the stitch (it makes the untrained MoE reproduce one expert exactly), but it is the wrong starting point for training: a perfectly symmetric gate can only be broken by the load-balancing loss, and three router trainings on one zero-init skeleton each collapsed onto a single expert with a different winner each time. That is why random is the default.

corpus: — how much text, and how varied

Two things here talk to each other, and you should know it before it surprises you:

• min_samples rises to meet router_mix_total. The mix is drawn from the .train split only — held-out has to stay held out — so each expert needs roughly router_mix_total ÷ experts ÷ 0.9 collected documents before the gate can be fed. The floor you set is a minimum; if the mix needs more, the floor goes up and the run tells you so: [cfg] corpus floor raised to 1,556 docs/expert …. It never goes down.
• max_samples below what the mix needs is refused at validate, because otherwise the corpus stage passes, every specialist trains, and the router comes up short of quota hours later — which reads as a gate that wouldn't learn when the truth is a gate that wasn't fed.

per_repo_cap is not a tuning knob, it's a correctness one. Measured: a C# bucket filled 78% of its token quota out of a single enterprise codebase. The resulting expert was fluent, passed every downstream check, and had learned one company's house style rather than the language. Lower is more diverse and needs more shards; if you raise it, you are trading variety for a shorter scan and you should mean it.

eval: — how the result gets measured

dead_threshold: 1.2 is deliberately above what a 150-step router produces. That is the point: an under-trained gate should be reported as undiscriminating, not quietly passed.

gates: — where the build stops and asks

main_evals: auto removes the last human checkpoint. It will happily run the full suite against a NaN'd model that generates at full speed and emits one token forever.

runtime: — the box, not the model

smoke: and roots: — the small print

Keep the {size} in roots.output. Without it every rung of the ladder writes to the same directory and the 3B run quietly eats the 0.5B one.

If you only turn three knobs

1. budget.target_steps — how good each specialist gets, and most of your wall-clock.
2. router.epochs — how opinionated the gate gets, at close to zero cost. If your eval says undiscriminating, this is the knob, not the corpus.
3. corpus.per_repo_cap — whether your expert learned a language or one codebase.

Everything else is refinement.

Using a template

Templates fill in name, base model, expert list, budget, and MoE config so you don't have to:

template: dnd

experts:
  - name: monster_manual
    source: { kind: hf, repo: PleiaSys/DnD-MonsterManual, text_field: text }
  - name: players_handbook
    source: { kind: hf, repo: PleiaSys/DnD-PlayersHandbook, text_field: text }
  - name: dm_guide
    source: { kind: hf, repo: PleiaSys/DnD-DMG, text_field: text }

Available templates: code, dnd, math, culinary.

Source kinds

Kinds are a registry, not a fixed list — another package can publish its own via the ms_moe_maker.corpus_kinds entry point without sending a PR here.

gh fetches one tarball from codeload rather than cloning, so there is no git binary needed and no history downloaded. Globs are matched against paths relative to the repo root, and **/ means zero-or-more directories the way a shell means it:

  - name: llama_docs
    source: { kind: gh, repo: ggml-org/llama.cpp, glob: "docs/**/*.md" }
  - name: my_wiki
    source: { kind: gh, repo: me/notes, ref: main, subdir: wiki, glob: "**/*.md" }

Public repos only, deliberately: a recipe is a document people share, which makes it the wrong object to put a credential in.

See recipe.example.yaml for the fully annotated version.

The tools (MCP) expert

A tool-calling specialist is the largest domain contrast a Ms.MoE can have — chat-formatted JSON-RPC against raw source — and it is the one expert whose corpus has to be generated rather than scraped. So it gets a dedicated knob instead of a pile of synth plumbing:

tools_expert: true

That injects a default tools expert (named agentcore, kind: synth, a sensible default teacher) into your expert list — a recipe with two code experts becomes a three-expert MoE with no other changes. To customise it, give a mapping instead of true:

tools_expert:
  name: my_mcp        # what the specialist (and its directory) is called
  teacher: Qwen/Qwen2.5-7B-Instruct   # the model that generates the traces

Anything you set wins over the default; kind is always synth. If an expert of that name already exists in your experts: list, it is used as the tools expert rather than duplicated.

Reasoning bases, and forcing reasoning in

Two different things, two knobs.

base_kind — is the base model already a reasoner?

base_kind: auto            # auto | reasoning | nonreasoning

auto sniffs the model id against the known reasoning families (see below); set it explicitly when the id isn't a known reasoning name. It only changes how the pipeline formats prompts and how eval reads the output — it does not make a non-reasoning base reason.

reasoning: true on a source — bake reasoning into a specialist that doesn't have it. This is the R1-distill recipe: a reasoning teacher writes <think>…</think> + answer pairs on the expert's domain, and the specialist is fine-tuned on them. It works on any base, including a small non-reasoning Qwen:

experts:
  - name: python
    source: { kind: stack, language: Python, reasoning: true }

The default reasoning teacher is deepseek-ai/DeepSeek-R1-Distill-Qwen-7B (-1.5B on a dryrun); set teacher: on the source to override it.

The tag table is a file, on purpose

The families that ship: DeepSeek, Qwen, OpenThink → Standard XML (<think>…</think>); Kimi → Interleaved Agentic XML; Llama → system-header tags. Eval scores the answer, and reports separately how often the model actually emitted a think block.

A wrong tag style is a silent wrong answer, not a crash. The splitter finds no delimiters, reports "did not reason", and the whole think block gets scored as if it were the answer. So when a new model family ships a new delimiter, you should not have to wait for a release — drop a file:

# ~/.msmoe/reasoning.yaml   (or point $MSMOE_REASONING at one)
Families:
  - Key: acme
    FamilyName: Acme Thinkers
    Models: [Acme-R2, acme-thinker]     # write what's on the model card
    PreferredStyle: xml

Layers merge by name — adding one family never costs you the other four — and model names match loosely: case, spaces, dots and hyphens are ignored on both sides, so Llama 3.1 matches meta-llama/Llama-3.1-8B-Instruct. The longest matching name wins, so the answer never depends on the order of the file.

And because the two cases look identical from the outside, eval says so rather than guessing:

almost nothing emitted a think block, and this run expected '<think>'…'</think>'.
That is either a model that does not reason or the WRONG TAG STYLE - the two
look identical from here, and every quality score above includes the trace if
it is the second.

The tags a run writes are stamped into its resolved config and carried, so eval splits with exactly the delimiters the generator used — and editing your table correctly changes the build_id.

Where llama.cpp lives

The GGUF export shells out to convert_hf_to_gguf.py, which lives in a llama.cpp checkout rather than on PyPI. That path is the most box-specific thing in a build, so a recipe can carry it:

runtime:
  llama_cpp: /mnt/nvme/llama.cpp

Resolution order is recipe → MSMOE_LLAMA_CPP → a short search of ./llama.cpp, ../llama.cpp, ~/llama.cpp and /opt/llama.cpp. The search looks for the converter itself, not just a directory with the right name.

Not finding it is a warning, never a failure: you still get the HF checkpoint, which is a real result. export and smoke are also the only two verbs that need no ML stack at all — see the install table above.

Preflight

Every build starts by asking the cheap questions, so the expensive part never starts on a box that cannot finish it:

• is torch / transformers / safetensors installed?
• is the base model reachable (or is it gated, or a typo)?
• are the roots writable, with enough room for the specialists + the stitched MoE + a GGUF?
• do the local corpus paths exist?
• is llama.cpp present? — a warning, not a failure. Without it you still get the HF checkpoint; you just do not get a GGUF.

ms-moe-maker build recipe.yaml --plan runs the same checks and stops there.

Failures carry their remedy, because the person reading one is usually about to lose an evening.

The pipeline

A recipe flows through six stages:

1. data.corpus — Collect expert corpora (stack scan, HF download, local files)
2. data.synth — Generate synthetic traces (if kind: synth experts)
3. finetune.{expert} — LoRA specialist training (one stage per expert)
4. stitch — Assemble the MoE skeleton from specialist checkpoints
5. router — Train the router gate weights (stratified expert mix)
6. export.gguf — Export GGUF and smoke-test it

The pipeline is fully modular. Each stage is an independent Python module. The orchestrator (builder.py) runs them in order, reports progress via a callback, and resumes from where it left off on re-run.

Evaluation

After a build, you can check whether your experts actually diverged:

ms-moe-maker eval recipe.yaml

Three questions, separately runnable:

ms-moe-maker eval recipe.yaml --mode routing   # the dead-expert check
ms-moe-maker eval recipe.yaml --mode quality   # generation vs held-out refs
ms-moe-maker eval recipe.yaml --mode experts   # did the specialists diverge at all?
ms-moe-maker eval recipe.yaml                  # all of the above (default)

Routing is the one Ms.MoE uniquely claims, and it is why hand-assigned experts work at all. A dead expert is not one that writes badly — it is one the router never routes to. So the measurement is routing, not text quality: held-out text from each expert's own domain goes through the MoE, the gate decisions are captured, and each expert gets an enrichment score — how much more it is used on its own domain than on average. Above ~1.2x means the router can tell that domain apart. Around 1.0x means it cannot, and that expert is dead however well it generates.

The report also names which expert is eating a weak one's traffic, because an expert can clear the enrichment bar and still be outranked on its own domain by a neighbour. That is a different failure, and a column-only read misses it.

Quality is real generation against held-out references. It needs an answer key, and whoever wrote the corpus is the only one who has it — which is exactly why this half is meant to be overridden.

Exit codes

3 exists on purpose. "We could not measure it" must never share an exit code with "it passed."

Overriding it

We provide the floor. Both halves are yours to replace, from the recipe:

eval:
  script: my_eval.py        # replaces ours entirely
  mode: routing             # routing | quality | experts | all
  held_out_fraction: 0.1
  num_samples: 20
  dead_threshold: 1.2       # minimum enrichment before "dead"

smoke:
  tokens: 48
  timeout: 300
  prompt: "Write a function that works."

A custom script is called as

my_eval.py --data-root R --output-root O --held-out F --num-samples N

which you can implement in any language you like.

Environment variables

The MSMOE_* names are inherited from the script this tool was carved out of and are still read for compatibility. New levers get MSMOE_*.

Supported base models

The fine-tune stage is generic — AutoModelForCausalLM will train a specialist from almost anything. The stitch stage is not: it builds a Qwen2MoeConfig, so today the base has to be a Qwen model.

validate refuses an unsupported base up front, deliberately. Without that check a Llama base collects its corpora, trains every specialist over several hours, and then dies at stage 4 — the most expensive possible place to find out.

Licence

GPL-3.0-only.