OmegaGPT SFT 145M
OmegaGPT SFT 145M is a small, experimental chat model from the OmegaGPT project. It uses the original legacy OmegaGPT decoder, a SentencePiece tokenizer, and native SafeTensors weights. It is intended for local experimentation, learning, and testing the architecture.
This is not a Llama/GGUF model or a standard transformers.AutoModel
checkpoint. Ollama and LM Studio do not load this release directly. The
lightweight Omega Chat runner below loads it without MIR Builder.
Quick start: Windows x64
- Download the chat runner files into one folder, keeping the
model/andfonts/subfolders intact. If the model weights are not beside the runner, the installer downloads them. - Double-click Install.bat and choose C for CPU or N for NVIDIA.
- When installation finishes, double-click Start Chat.bat.
- If your browser does not open, visit http://127.0.0.1:7860.
No existing Python installation or administrator access is needed. Python 3.12,
PyTorch, dependencies, and downloaded weights stay under .runtime/ beside
the runner. First setup needs internet and several gigabytes of free disk
space. NVIDIA mode checks whether PyTorch can actually use your GPU; if that
fails, rerun Install.bat and choose CPU. Rerunning setup reuses verified files.
The chat server listens only on your computer. Do not upload .runtime/ back
to Hugging Face: it contains generated dependencies and caches.
Manual / other platforms
Install a suitable PyTorch build and then run:
python -m pip install -r requirements.txt
python chat.py
If weights are not next to chat.py, first launch downloads them from
Ried-wqs/omega-sft-145m.
For terminal use:
python chat.py --cli
python chat.py --prompt "Hello!"
python chat.py --device cpu --model path/to/export
Use --device cuda for a supported NVIDIA/PyTorch setup; the default
--device auto selects CUDA when available. Use --max-tokens 80 to lower
generation time or --port 7861 if the default port is occupied.
The interface bundles IBM Plex Sans Condensed under its SIL Open Font License;
see fonts/OFL.txt.
What is in this release?
| Item | Purpose |
|---|---|
model.safetensors | Trained model weights |
config.json | OmegaGPT architecture settings |
tokenizer.model | SentencePiece tokenizer |
manifest.json | Optional checksum verification |
chat.py, config.py, model/ | Standalone loader and chat application |
index.html, fonts/ | Local browser interface and font |
Install.bat, Start Chat.bat, setup.ps1 | Windows installer and launcher |
Format and limitations
The repository must include model.safetensors, config.json, and
tokenizer.model; manifest.json enables checksum verification. The runner
supports the legacy OmegaGPT export represented by this model, not the newer
Omega2 or Next engines. Its browser UI is local-only and has no user accounts.
This is a small experimental model, not a production assistant. Outputs may be inaccurate, repetitive, or inconsistent. Do not rely on them for medical, legal, financial, or other consequential decisions. Exact training data provenance, benchmark results, and the model's redistribution license should be added after verification; they are intentionally not claimed here.