Recommended Usage
For the easiest way to use this model with all required settings and its full functionality, use the TrajectoryLM application.
Screenshots
SmolLM2-135M V3 — edge/travel trigger
This checkpoint freezes SmolLM2 and trains a 1,328,257-parameter branch that measures changes between consecutive hidden rows (edges), accumulates their magnitudes (travel), and uses that trajectory to adjust selected following token distributions. Runtime trigger words and optional context are controlled by the companion comparison UI.
| Property | Value |
|---|---|
| Unique parameters | 135,843,265 |
| Trained branch parameters | 1,328,257 |
| Layers / width | 30 / 576 |
| Context length | 8,192 |
| Completed branch updates | 2,000 |
| Token presentations | 262,144,000 |
| Terminal sampled validation loss | 1.1849 (PPL 3.3) |
The recorded validation value is a terminal ten-batch training-loop estimate for this branch objective, not a standardized full-stream language benchmark.
Local UI
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 checks for local weights and can download the Hugging Face release
with visible progress. See research_paper.pdf for equations, provenance and
limitations. The repository-native PyTorch implementation is required.
Limitations
This is an architecture experiment, not a production assistant. Guidance can be too weak to redirect text or strong enough to damage fluency and cause loops. It has no comprehensive capability or safety evaluation.
Citation
@misc{argo1ooas2026smollm2v3,
title={Contextual Edge/Travel Trigger Adaptation of SmolLM2-135M},
author={Argo1-OOAS}, year={2026},
url={https://huggingface.co/Argo1-OOAS/SmolLM2-135M-V3-EdgeTravel}
}