Modelfile

A Modelfile describes a derived model: a base model plus the questions it should always ask, and optionally a refit calibration, a pinned precision, a license and a description. Build it with ollaya create, then run it by name with just a state.

ollaya create triage -f Modelfile
ollaya run triage "I was charged twice for my subscription this month."

Example

# Support ticket triage
FROM laya:en

QUESTIONS """
{
  "department": {
    "type": "choice",
    "instructions": "Which team should handle this ticket?",
    "criteria": {
      "billing": "Payments, invoices and refunds",
      "technical": "Bugs, errors and outages",
      "account": "Login, profile and settings"
    }
  },
  "urgency": {
    "type": "score",
    "instructions": "How urgent is this ticket?",
    "criteria": ["Can wait", "Needs attention this week", "Needs attention today"]
  },
  "refund": {
    "type": "noul",
    "instructions": "Does the customer ask for money back?"
  }
}
"""

CALIBRATION ./calibration.json
PARAMETER precision fp32
DESCRIPTION Support ticket triage
LICENSE ./LICENSE

Instructions

Instruction	Required	Value
`FROM`	yes	The base model, e.g. `laya:en`, or a router such as `laya`
`QUESTIONS`	no	The built-in questions: the same object as `questions` in the API
`CALIBRATION`	no	Temperatures that replace the base model's
`PARAMETER`	no	`precision fp16` or `precision fp32`, to pin one graph
`DESCRIPTION`	no	One line, shown by `ollaya show` and `/v1/models`
`LICENSE`	no	License text for the derived model

Syntax. Directives are case-insensitive; # starts a comment line.
Values. QUESTIONS, CALIBRATION and LICENSE take a path relative to the Modelfile (~/ works), inline JSON on one line, or a block between """ spanning several lines.
Unknown input. Any other directive or parameter is an error.
Pulling. ollaya create pulls FROM first when it is not on this machine.

Anything you leave out is inherited from the base model. When a request sends its own questions, they replace the built-in ones for that request.

FROM

The model to build on. A router works too; the derived model then routes like its base:

FROM laya

QUESTIONS

The questions the model asks when a request brings none, validated like a decision request (1–256 questions):

QUESTIONS ./questions.json

CALIBRATION

Probabilities are only useful for thresholds if they are calibrated. Ollaya calibrates with temperature scaling: logits are divided by a temperature per question type and number of options. CALIBRATION replaces the base model's temperatures with ones refit on your own labelled data:

{
  "temperature": [1.6, 1.25, 1.98],
  "temperature_by_options": {
    "choice:2": 1.9,
    "choice:3-5": 1.76,
    "score:3-5": 1.25,
    "noul:2": 1.98
  }
}

temperature holds one fallback per question type (choice, score, noul). temperature_by_options keys are <type>:<2|3-5|6-10|11+>, by the question's number of options.

PARAMETER

The only parameter is precision. A bare model carries an fp16 and an fp32 graph and picks one when it loads (fp16 on a CUDA GPU, fp32 on the CPU). PARAMETER precision fp32 pins the fp32 graph everywhere, for example to match an fp32 reference exactly.

PARAMETER precision fp32

DESCRIPTION

DESCRIPTION Support ticket triage

LICENSE

The license text shipped with the model and shown by ollaya show --license. When you build on an Apache-2.0 model such as Laya, keep its license and attribution.