Recommended Usage
For the easiest way to use this model with all required settings and its full functionality, use the TrajectoryLM application.
Screenshots
TinyLlama-1.1B — Argo1-OOV1
Argo1-OOV1 adds a learned goal-conditioned path to a frozen TinyLlama chat
Transformer. It measures the prompt left-to-right, measures a user-supplied end
goal right-to-left, fuses their edge/travel mismatch into a bounded hidden
residual, and can retain beam routes with the lowest accumulated
-log(probability) energy.
| Property | Value |
|---|---|
| Total parameters | 1,121,021,953 |
| Trained goal-branch parameters | 20,973,569 |
| Layers / width | 22 / 2,048 |
| Context length | 2,048 |
| Completed branch updates | 2,000 |
| Token presentations | 262,144,000 |
| Terminal sampled validation loss | 1.0901 (PPL 3.0) |
This loss measures the synthetic future-goal branch objective. It does not show that generated answers reach arbitrary human goals or outperform TinyLlama.
Use
git clone https://github.com/Argo1-OOAS/TrajectoryLM.git
cd TrajectoryLM
Run start_windows.ps1 on Windows or start_macos.command on macOS, following
the TrajectoryLM README.
The launcher detects missing weights, downloads them with progress, and shows
research_paper.pdf. The UI exposes the end goal, guidance, beam width, token limit,
lowest-energy route and per-token normal-versus-goal choice audit.
Limitations
Goal guidance may be too weak, may trade fluency for lexical overlap, or may take a long time because the current beam audit recomputes prefixes without a KV cache. “Lowest energy” means highest model probability under this learned goal-conditioned distribution; it is not physical energy and does not prove correctness, safety, or plan optimality.
Citation
@misc{argo1ooas2026argo1oov1,
title={Argo1-OOV1: Goal-Conditioned Bidirectional Edge/Travel Steering},
author={Argo1-OOAS}, year={2026},
url={https://huggingface.co/Argo1-OOAS/TinyLlama-1.1B-Argo1-OOV1}
}