--- title: Pidgin Test emoji: 💻 colorFrom: gray colorTo: purple sdk: gradio sdk_version: 6.5.1 python_version: '3.12' app_file: app.py pinned: false --- # Pidgin Test ## Overview This is a small test Space for trying out a LoRA adapter that fine-tunes [`openai/gpt-oss-20b`](https://huggingface.co/openai/gpt-oss-20b) to reply in Nigerian Pidgin English. It loads the base model, applies the [`AnalyticsIntelligence/pidgin_oss`](https://huggingface.co/AnalyticsIntelligence/pidgin_oss) adapter with PEFT, and exposes a simple Gradio chat interface (with sliders for max tokens, temperature, and top-p) so the adapter's behavior can be checked interactively. It is a GPU-only inference demo/experiment rather than a finished product — there is no evaluation harness or benchmarking, just a chat UI for manually probing how the fine-tuned model responds in Pidgin. ## Tech Stack - **Base model**: `openai/gpt-oss-20b` - **Adapter**: `AnalyticsIntelligence/pidgin_oss` (LoRA, applied via PEFT) - **Libraries**: `transformers`, `peft`, `accelerate`, `bitsandbytes`, `torch` - **Quantization**: MXFP4 (`Mxfp4Config` with `dequantize=True`) for running on non-H100 GPUs (L4/A10/T4, etc.) - **UI**: Gradio `ChatInterface` ## Author Built by [Ephraimmm](https://huggingface.co/Ephraimmm)