---
license: apache-2.0
language: [en]
library_name: pytorch
pipeline_tag: text-generation
base_model: HuggingFaceTB/SmolLM2-135M-Instruct
datasets: [HuggingFaceTB/smol-smoltalk]
tags: [causal-lm, research, quadorbit, trigger-conditioned, custom-code]
---


## Recommended Usage

For the easiest way to use this model with all required settings and its full functionality, use the [TrajectoryLM application](https://github.com/Argo1-OOAS/TrajectoryLM).

## Screenshots

![TrajectoryLM interface](./Screenshot%202026-08-26%20185138.png)

![TrajectoryLM model view](./Screenshot%202026-08-26%20203341.png)

# SmolLM2-135M V2 — QuadOrbit trigger

This research checkpoint keeps the SmolLM2 Transformer core frozen and trains a
424,641-parameter QuadOrbit/context branch. At inference, a user-selected
trigger activates the branch and can alter a configurable number of following
token distributions. Optional context conditions that private branch; it does
not modify weights during prompting.

| Property | Value |
|---|---:|
| Unique parameters | 134,939,649 |
| Trained branch parameters | 424,641 |
| Layers / width | 30 / 576 |
| Context length | 8,192 |
| Vocabulary | 49,152 |
| Completed branch updates | 2,000 |
| Token presentations | 262,144,000 |
| Terminal sampled validation loss | 1.6958 (PPL 5.5) |

The validation number is the training loop's final ten-random-batch estimate on
the local SmolTalk validation stream. It is not a standardized benchmark and is
not directly comparable to another objective or tokenizer.

## Use in the comparison UI

Use this checkpoint through the TrajectoryLM comparison application:

```bash
git clone https://github.com/Argo1-OOAS/TrajectoryLM.git
cd TrajectoryLM
```

Then run `start_windows.ps1` on Windows or `start_macos.command` on macOS, as
described in the [TrajectoryLM README](https://github.com/Argo1-OOAS/TrajectoryLM#readme).
The interface checks for the release, offers a download if it is absent, shows
download progress, and exposes the trigger/context controls and token audit.
The release uses repository-native PyTorch code rather than Transformers
`AutoModel`; review the source before execution.

## Paper and limitations

Open `research_paper.pdf` for the architecture and training report. This is an
experimental branch, not a reliable assistant. It can repeat, hallucinate,
ignore added context, or lose fluency at high guidance. Do not rely on it for
medical, legal, financial, safety, or other high-impact decisions.

## Citation

```bibtex
@misc{argo1ooas2026smollm2v2,
  title={Trigger-Scoped QuadOrbit Adaptation of SmolLM2-135M},
  author={Argo1-OOAS},
  year={2026},
  url={https://huggingface.co/Argo1-OOAS/SmolLM2-135M-V2-QuadOrbit}
}
```