Spaces:
Running
Download README.md from valueai-org/README: direct link, hf CLI and curl.
- Browser
- Download file 2.58 kB
-
https://huggingface.co/spaces/valueai-org/README/resolve/main/README.md
- Command line
-
hf download hf://spaces/valueai-org/README/README.md
-
curl -L -o README.md https://huggingface.co/spaces/valueai-org/README/resolve/main/README.md
title: ValueAI
sdk: static
app_file: index.html
license: mit
ValueAI
valuealigned.ai shares model research, source code, and study artifacts that others can inspect and test.
Signal-Latent World Models
SLWM studies how a model can use current sensory signals, retain and revise working memory, and decide when its evidence supports an action. This work involves valuealigned.ai and keplen.ai.
The manuscript draft v0.3 is unpublished. The models below are study artifacts for reference, each from a fixed seed-2001 run. They are not formal model releases.
| Study artifact | What it represents |
|---|---|
| SLWM-X1B | Wave 5, 1.008B parameters: the larger world-model checkpoint with doubled training exposure. |
| SLWM-X450M-Memory | Wave 6, 453M parameters: the full model used as the memory and ablation reference. |
| SLWM-v2-453M | Wave 7, 453M parameters: the spatial-grid reference checkpoint. |
Each model repository contains the original BF16 and INT8 exports, configuration, run records, checksums, and a loader. The recorded metrics in each repository belong to one training seed. The manuscript's comparative tables summarize three seeds per cohort and use different evaluation settings across waves.
Start with the runnable examples and validation-set guide. The guide shows a small generated example, the required recorded-data banks and their split names, and commands for development and held-out evaluations. The recorded media and robot banks are not included in the model repositories.
Sources and license
- Expanded scientific article, figures, limitations, and next steps
- SLWM source repository
- MIT license and copyright notice
The study artifacts use the same MIT license as the GitHub project. The copyright owners are Diogo Antunes (contact@diogoantunes.ai), ValueAI (contact@valuealigned.ai), and Keplen (contact@keplen.ai).