File size: 2,577 Bytes
dc7f752
0b13228
dc7f752
0b13228
 
dc7f752
 
0b13228
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
---
title: ValueAI
sdk: static
app_file: index.html
license: mit
---

# ValueAI

[valuealigned.ai](https://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](https://valuealigned.ai) and [keplen.ai](https://keplen.ai).

The [manuscript draft v0.3](https://diogoantunes.ai/media/publications/slwm_x1b_paper_v03.pdf) is **unpublished**.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](https://huggingface.co/valueai-org/SLWM-X1B) | Wave 5, 1.008B parameters: the larger world-model checkpoint with doubled training exposure. |
| [SLWM-X450M-Memory](https://huggingface.co/valueai-org/SLWM-X450M-Memory) | Wave 6, 453M parameters: the full model used as the memory and ablation reference. |
| [SLWM-v2-453M](https://huggingface.co/valueai-org/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](https://github.com/ValueAlignedAI/slwm/blob/main/docs/research/x1b_study_artifacts_and_validation.md). 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](https://diogoantunes.ai/publications/signal-latent-world-models)
- [SLWM source repository](https://github.com/ValueAlignedAI/slwm)
- [MIT license and copyright notice](https://github.com/ValueAlignedAI/slwm/blob/main/LICENSE)

The study artifacts use the same MIT license as the GitHub project. The copyright owners are **Diogo Antunes** ([contact@diogoantunes.ai](mailto:contact@diogoantunes.ai)), **ValueAI** ([contact@valuealigned.ai](mailto:contact@valuealigned.ai)), and **Keplen** ([contact@keplen.ai](mailto:contact@keplen.ai)).