OpenWALDO

Models with receipts

Train the model.
Keep the evidence.

02 / TRAINING

WALDO resolves verified corpus data before execution, records the starting point and observed run, and binds the resulting weights into inspectable release BOMs.

Explore the WALDO toolchain ↗

Two honest starting points

Begin blank.
Or begin with lineage.

WALDO does not blur a fresh architecture and an acquired checkpoint. Each enters the lifecycle with different evidence.

START / 01Blank architecture

Declare the model.

Use a strict compose to define the architecture, ordered corpora, objectives, and training parameters. WALDO creates and trains the model from that portable plan.

Trainwaldo model train small ./pretrain.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.

Pull compatible Hugging Face Safetensors at an immutable revision, validate them, and record an origin BOM. A compose with base.model: llama-base can then create a new model, adding training without mutating the original.

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

The working lifecycle

Plan before allocation.
Record after reality.

  1. 01
    RESOLVEVerified corpus BOM

    Pin index revision, manifests, sources, licenses, shards, counts, and hashes.

  2. 02
    FORECASTFit before you run

    Estimate memory and runtime across exact supported Apple, NVIDIA, and AMD profiles.

  3. 03
    TRAINExecute and observe

    Run on MLX, PyTorch, or single-node TorchTitan; persist environment, consumption, losses, checkpoints, and output hashes.

  4. 04
    VALIDATEInspect the result

    Evaluate on a deterministic held-out selection and test generation where a compatible runtime exists.

  5. 05
    EXPORTRelease one representation

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

The provenance package

Not “trust me.”
Here is the record.

Each artifact has a specific job. Together they connect selected data, a model starting point, execution, and release inventory without pretending a hash proves more than identity.

DATA

Corpus BOM

The exact resolved index selection.

ORIGIN

ORIGIN-BOM.json

The immutable acquired checkpoint, when present.

RUN

RUN-BOM.json

The plan, environment, observations, and output identity.

RELEASE

BOM.json + EU-BOM.json

Technical inventory and regulatory disclosure projection.

Useful now. Narrow on purpose.

Working end to end,
with visible boundaries.

Implemented today

  • Real MLX training and generation on Apple Silicon
  • PyTorch on Linux CPU, NVIDIA CUDA, and AMD ROCm
  • Single-node distributed TorchTitan
  • Verified evaluation, resume, export, and optional signing

Deliberately pending

  • General Hugging Face tokenizer and architecture support
  • SFT, preference training, and pinned chat templates
  • PyTorch generation and multi-node orchestration
  • Supported public packaging and releases
Explore the WALDO toolchain ↗Get help from the community ↗