Orpheus TTS — PT-BR LoRA Fine-tuning (Unsloth)

LoRA fine-tuning of Orpheus 3B (canopylabs/3b-es_it-ft-research_release) via Unsloth, run artoodtoo_ft_e3. Training continued from the PT-BR adapter ArtooDtoo/Orpheus_PTBR_FT_Unsloth_Chk-1800. Adapters and reports at the root; graficos/ with curves and timing; inferencias/ with evaluation audio.

Checkpoints (LoRA adapters)

Path	Description
`final_adapter/`	Final adapter (372 MB) — use this for inference
`checkpoint-3351/`	Adapter at step 3,351 + tokenizer + trainer state

Training

Parameter	Value
Base model	canopylabs/3b-es_it-ft-research_release
Starting adapter	ArtooDtoo/Orpheus_PTBR_FT_Unsloth_Chk-1800
Train / validation samples	13,404 / 273
Steps	3,351 (1 epoch)
Training time	~~21.4 h (~~23 s/step)

Reports in training_report.txt and training_stats.json.

Files

graficos/ — loss_curve*.png, losses.csv, epoch_times.csv, step_times.csv
inferencias/ — artoodtoo_ft_e3_epoch1_p*.wav (end of epoch 1), artoodtoo_ft_e3_{custom3,random3,retry_clean,retry_clean2}_p*.wav (evaluation in 4 configurations), and the prompts used (*prompts*.txt)

Usage

Load the base model and apply the adapter with PEFT/Unsloth:

from peft import PeftModel
model = PeftModel.from_pretrained(base_model, "OrpheusTTS/final_adapter")

Citation

These adapters are LoRA fine-tunes of canopylabs/3b-es_it-ft-research_release, continued from ArtooDtoo/Orpheus_PTBR_FT_Unsloth_Chk-1800. If you use them, please cite:

DINIZ, Thomaz. Incorporando regionalismos e sotaques em modelos de síntese de fala. 2026. Dissertação (Mestrado em Ciência da Computação) — Universidade Federal de Campina Grande, Campina Grande, 2026.