# Orpheus-3B pt-BR (LoRA adapter)
A LoRA adapter that teaches [Orpheus-3B 0.1 (pretrained)](https://huggingface.co/unsloth/orpheus-3b-0.1-pretrained) to speak **Brazilian Portuguese**. Orpheus is a Llama-based text-to-speech model that generates speech as a sequence of [SNAC](https://github.com/hubertsiuzdak/snac) audio tokens. This adapter was trained from the English-centric *pretrained* checkpoint (not an already fine-tuned voice model) on Portuguese speech data, using [Unsloth](https://github.com/unslothai/unsloth) and two NVIDIA T4 GPUs on Kaggle.
> This repository contains **only the LoRA adapter weights**. You need the base model `unsloth/orpheus-3b-0.1-pretrained` and the SNAC 24 kHz codec (`hubertsiuzdak/snac_24khz`) to generate audio.
## Model summary
| | |
|---|---|
| Base model | `unsloth/orpheus-3b-0.1-pretrained` (3B parameters, Llama architecture) |
| Method | LoRA (r = 64, alpha = 64) on all attention and MLP projections, base weights kept in 16-bit (no quantization…
