ChudGPT-Public

Author: Astra228. Small experimental language model: 20,999,184 parameters. Custom PyTorch implementation, SafeTensors weights, and original tokenizer. This is a downloadable model, not a hosted chat service.

Checkpoint versus the live application

This repository exports the selected public_assistant_rebalance/latest.pt neural checkpoint. It is a custom, independently trained decoder-only model, not a fine-tune of an external model. The live Public V20 application combines neural generation with additional reviewed responses, exact arithmetic, normalization and other local answer-handling logic. Those application layers are not reproduced by this minimal neural-only runner, so live-bot behavior and application test results must not be treated as results for these weights alone. The separate Music and LGBTQ specialist checkpoints are not included.

This upload does not claim standardized benchmark scores or ChatGPT-level quality. Expect incorrect answers, hallucination, repetition, weak reasoning and limited conversation memory. The architecture and prompt-formatting code are included for reproducible local inference.

Run locally

git clone https://huggingface.co/Astra228/ChudGPT-Public
cd ChudGPT-Public
python -m pip install -r requirements.txt
python run.py --message "Hello!" --device cpu
python run.py --device cuda

Install Git LFS before cloning to download the weights, or download all files from the Hub. CUDA requires a compatible GPU and PyTorch build. Interactive history is kept only in the current process. /clear resets it; /quit exits. Default decoding is deterministic. These files use custom local Python code, not transformers.AutoModel or an Inference Provider. Review code before running it. No service credentials, Discord integration, private chat logs, training corpus or optimizer/resume state is included.

Provenance and limitations

The original local checkpoint SHA256 is f551a6cb7141aa840141b5946fa56bdfdd73409c8ee24c5a4396a056c5a61320. Export is tensor-exact relative to that checkpoint; config.json records architecture and step. The original checkpoint is preserved locally. Training-data provenance is not a claim of an exhaustive third-party rights audit; the raw training datasets are not redistributed here. Do not rely on output for consequential decisions. Check generated code and math independently.

License

The source project license notice is included in LICENSE (Apache-2.0).