📜 Docs

Quick start for the distillery CLI. Full reference coming soon.

Install

pip install distillarium[gemini]
distillery --version
# → distillery, version 0.1.0

Set a teacher API key

# Gemini Flash is the default teacher (cheapest at ~$0.30 / 1000 examples).
export GOOGLE_API_KEY=your-key-here

Distill your first Spirit

# Use the reference Needle recipe (15 tool categories, ~20.7M params).
distillery distill recipes/needle.tool-calling-v1.yaml --out spirits/

# Expected: ~30 minutes, ~$0.30 in Gemini API, 78° proof on held-out.

Inspect your Cellar

distillery cellar
#                🍾 Your Cellar
# ┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━┓
# ┃ Spirit              ┃ Proof ┃ Params ┃
# ┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━┩
# │ needle.tool-calling │  78°  │ 20.7M  │
# └─────────────────────┴───────┴────────┘

Re-taste against fresh data

distillery taste spirits/needle.pt --mash new_held_out.jsonl

Bottle for deployment

# ONNX for cross-runtime inference
distillery bottle spirits/needle.pt --format onnx

# GGUF for llama.cpp / mobile (coming in v0.2)
distillery bottle spirits/needle.pt --format gguf

Recipe schema

name: needle.tool-calling
version: 1

teacher:
  provider: gemini
  model: gemini-2.5-flash
  temperature: 0.9

mash:
  total_examples: 1000
  examples_per_call: 10

student:
  arch: attention-only-glu
  d_model: 384
  n_heads: 6
  n_layers: 8

cuts: { train: 0.9, eval: 0.1 }

still:
  epochs: 8
  batch_size: 16
  lr: 3.0e-4

tasting:
  metrics: [tool_name_accuracy, arg_key_f1, exact_call_accuracy]
  held_out: 100