OpenWALDO

Models begin as composes

Compose the model.
Keep the evidence.

02 / TRAINING

WALDO treats the model compose as the source definition for a reproducible model build. Forecast it, run it, resume it, inspect it, and export the resulting model with its lineage attached.

Read the model compose guide ↗

The model build definition

One compose.
The whole plan.

A model compose is strict, portable YAML or JSON that defines what to build. It keeps model intent reviewable while WALDO chooses compatible machine-local execution for the hardware where it runs.

01 / MODEL

Architecture and interaction

Model shape, tokenizer, context, data type, and conversational contract.

02 / ORIGIN

Fresh or verified weights

Initialize from scratch, a managed model, or a pinned supported source.

03 / DATA

Exact corpus mixture

Named corpora, record filters, license exclusions, and relative weights.

04 / STAGES

Ordered model development

Pretraining, midtraining, conversation, and tool-use stages in sequence.

05 / CONTROL

Budgets and evaluation

Tokens or epochs, profiles, seeds, checkpoints, and held-out evaluation.

PORTABLE INTENT Describe the model once. Run it where it fits.

A compose deliberately excludes GPUs, credentials, framework paths, and other local policy. WALDO resolves a compatible backend, records what actually ran, and rejects unsupported or incompatible plans before training.

NEED A STARTING POINT?

Build a compose with the WALDO advisor.

$ waldo advisor <model-name> turns your goals, available data, hardware, and budget into a validated compose proposal. You review the plan; WALDO writes or trains it only after you approve.

Why composes matter

Reviewable before training.
Reproducible after.

A compose turns model building from a collection of commands and local assumptions into one durable artifact that people can inspect, discuss, run, compare, and improve together.

01 / REVIEW

Inspect the plan before spending compute.

Review the model architecture, origin, data mixture, filters, objectives, budgets, and evaluation policy as one coherent proposal.

02 / PORTABILITY

Separate model intent from the machine.

Keep hardware, credentials, paths, and backend policy local while the compose remains portable across compatible execution environments.

03 / CONTINUITY

Make a multi-stage build durable.

Execute stages in declared order, stop safely on failure, and resume an interrupted build from its latest verified checkpoint.

04 / LINEAGE

Extend models without losing their history.

Pin a verified base, preserve each compose and run, and create new model lineages without mutating the model that came before.

One format. Two starting points.

Begin blank.
Or begin with lineage.

WALDO does not blur a fresh architecture and an acquired checkpoint. Both use a compose, but each enters the lifecycle with different evidence and validation rules.

START / 01Blank architecture

Declare the model.

Omit the base and declare the complete architecture. WALDO initializes new weights, executes every ordered stage, and preserves the compose with the model history.

Forecastwaldo model forecast ./small.compose.yaml
Trainwaldo model train small ./small.compose.yaml
Listwaldo model list
Exportwaldo model export small ./small-hf --format huggingface
vLLMvllm serve ./small-hf
START / 02Supported open weights
SAFETENSORSORIGINSHA-256

Pin the origin.

Use base.model for a verified managed checkpoint or base.source for a supported external model pinned to an immutable revision. The compose creates a new lineage without mutating its base.

Pullwaldo model pull llama-base huggingface://organization/repository@revision
Forecastwaldo model forecast ./continued.compose.yaml
Trainwaldo model train llama-continued ./continued.compose.yaml
Listwaldo model list
Exportwaldo model export llama-continued ./llama-hf --format huggingface

The working lifecycle

Plan before allocation.
Record after reality.

  1. 01
    COMPOSEDeclare the build

    Define the origin, architecture, interaction, stages, data, and training budget.

  2. 02
    FORECASTFit before allocation

    Validate the compose and estimate parameters, memory, steps, and runtime without changing the model.

  3. 03
    TRAINResolve, run, and resume

    Pin every selected corpus and origin, execute stages in order, and continue an interrupted compose from its latest verified checkpoint.

  4. 04
    INSPECTCompare plan and reality

    Review the saved compose, run history, actual consumption, telemetry, held-out evaluation, checkpoints, and artifact hashes.

  5. 05
    EXPORTRelease with the record

    Package WALDO, Hugging Face, MLX, GGUF, or Ollama formats with technical and EU BOMs.

The AI Bill of Materials

Know what built it.
Audit every layer.

An AI Bill of Materials is only meaningful when it points back to the exact corpora behind a model. OpenWALDO makes those corpora versioned, addressable, verifiable, and auditable, then connects them to the model’s origin, compose, training run, and release.

01 / CORPUS

Corpus BOM

The exact, versioned corpus selection and its content-addressed records.

02 / ORIGIN

ORIGIN-BOM.json

The immutable identity and source of any model used as a starting point.

03 / BUILD

RUN-BOM.json

The compose, resolved inputs, environment, observations, and resulting artifacts.

04 / RELEASE

BOM.json + EU-BOM.json

The model’s portable technical inventory and regulatory disclosure record.