---
license: apache-2.0
language: en
tags:
  - waldo
  - experimental
  - base-model
  - safetensors
---
# WALDO 1B initial experimental reference checkpoint

**Experimental reference weights — Apache License 2.0.**

This is a raw causal base model, provided for reproducing WALDO training and
evaluation experiments. The frozen behavioral qualification did not pass.
It is not instruction-tuned, and character/topic drift and factual errors remain.
See `evaluation/` for the complete assessment and limitations.

## Exact artifact

- Managed model: `general-base-1b-a-tuned-01`
- Native run: `5652ba3bce3a7f6c`
- Selected checkpoint within stage: 271073
- Selected artifact: 4138071392 bytes, SHA-256 `f3159fa6b8be2acd8efbc0d9daebeba59dcaf2f1f9ed47ff925f656a187028cb`
- Held-out FP32-inference loss: 2.195388929
- Held-out bits per byte: 0.799569705
- Stage completed: 295716 steps / 4845010944 target tokens
- Selected snapshot token position within stage: 4441260032

The continued model starts from the initial model's selected trained weights,
with a new AdamW optimizer and a lower peak LR (1e-4 versus 4e-4). It replays
the same deterministic data stream; extra exposure is not unique new data.
The native BOMs retain the exact architecture, 32K byte-BPE tokenizer, corpus
mixture, held-out selection, runtime and training controls. No qualification
tolerances were relaxed.

## Download from Hugging Face

Download the complete repository, retaining its native/ directory structure:

```sh
hf download whiskeywhiskey/waldo-1b-initial-reference --local-dir ./waldo-1b-initial-reference
cd waldo-1b-initial-reference
```

Then use the WALDO commands below. The companion continued checkpoint is at
[waldo-1b-continued-reference](https://huggingface.co/whiskeywhiskey/waldo-1b-continued-reference).

## Loading and format

This is an **inference-only managed snapshot**, not a complete native archive
or a Transformers/GGUF conversion. It retains exact metadata and current
published artifacts; intermediate checkpoint/optimizer payloads are omitted
and enumerated in `PACKAGE.json`. The continued snapshot also retains its
required parent initializer under `base/`. It cannot restore every historical
checkpoint. The separate full-backup staging tree retains all source files.

Use a WALDO CLI with the original native PyTorch model/tokenizer support and
a compatible PyTorch environment. After downloading the repository:

```sh
waldo model summary "$PWD/native/general-base-1b-a-tuned-01"
CUDA_VISIBLE_DEVICES=0 waldo model chat "$PWD/native/general-base-1b-a-tuned-01" --raw --temperature 0 --seed 42 --max-tokens 64 "A map helps people"
```

The same snapshot is smoke-tested with CUDA hidden during preparation. It
does not require access to the original absolute training-data paths for
inference. Standard Hugging Face/GGUF export is currently unsupported for
this trained 32K tokenizer; do not assume `AutoModel.from_pretrained` works.

## Provenance, attribution and license

See `PACKAGE.json`, the immutable native model/run BOMs, the compose retained
in the managed metadata, and `ATTRIBUTION.md`. Training-data declarations
include CC-BY-SA-4.0, CDLA-Sharing-1.0 and ODC-BY-1.0 and remain preserved.
The operator explicitly chose Apache-2.0 for these weights to match WALDO's
project license. See LICENSE, NOTICE and LICENSING.md. This is an independent
experimental release, not an official OpenWALDO model.

The native records retain historical host/runtime paths for reproducibility.
Git commit signing and native model/export signing are separate; this package
does not claim a signed WALDO release or complete regulatory disclosure.