⌂ THE WORKBENCH · RAW SUBSTRATE

Distillarium — Workbench

The technical view. Same substance as the Specimen. No metaphors, no decoration — code, benchmarks, recipes, and notebooks. Open in Colab when you want to run something.

# install
pip install distillarium

# distill a teacher into a Spirit (uses recipes/needle.yaml)
distill recipes/needle.yaml --teacher claude --out spirits/needle

# evaluate
taste spirits/needle --eval bfcl

# bottle for deployment (ONNX export + adapter)
bottle spirits/needle --format onnx

Three commands cover the full pipeline: distill the teacher's behavior into a tiny student, taste-test against held-out cases, bottle for production.

Spirit	Task	Params	Headline metric	Source
Needle	Function / tool calling	~30M	78% tool-name accuracy on BFCL subset	Tasting Notes →
Routor	Intent routing	~25M	92% macro-F1 on internal eval	Tasting Notes →

Research code, not production infrastructure. Read each Spirit's full evaluation (strengths, weaknesses, failure cases) before deploying. Auto-generated Tasting Notes never hide the bad numbers.

name: needle
task: tool_calling
teacher:
  provider: anthropic
  model: claude-3-5-sonnet-20241022
student:
  arch: encoder_decoder
  params: 30_000_000
mash:                  # seed corpus
  size: 50_000
  domains: [api_calls, function_definitions, tool_use_examples]
cuts:                  # data splits
  train: 0.8
  val: 0.1
  proof: 0.1          # held-out, never seen
eval:
  benchmark: bfcl_subset

A recipe describes what to distill, from whom, and how to evaluate it. Full recipes for both bottled Spirits live in /recipes.

Walk through end-to-end distillation in a hosted notebook. Free CPU is enough for a small Spirit; bring your own teacher API key.

Open quickstart.ipynb on GitHub → Open in Colab →

The Colab link points at the canonical notebook in the repo — when the repo notebook updates, Colab pulls the latest on next open.

github.com/Lumi-node/distillarium — full source, MIT licensed
pypi.org/project/distillarium — installable package
/recipes — YAML configs for each Spirit
/spirits — bottled model weights and adapters
/scripts — validation runners (BFCL, teacher-vs-student)
Docs — full documentation index

Capability ceiling. Students can't exceed teacher quality on the distilled task. If your teacher is wrong on edge cases, your Spirit will be wrong in the same places.
Recipe quality dominates. Mash diversity matters more than student size. A 30M Spirit on a great mash beats a 50M Spirit on a thin one.
Out-of-distribution failure is silent. Spirits don't know what they don't know. Evaluate on real production traffic before you ship.
Not for everything. If you need flexible reasoning across many domains, keep a frontier teacher in the loop. Distillation works best for narrow, well-defined steps.