# Phase 2: LoRA SFT on Erinome dataset

**Date:** Mon Apr 27 03:22:08 UTC 2026
**GPU:** RTX 4090 inst-35658834 vast.ai
**Cost:** ~$1/hr × 1.5h ≈ $1.50

## Setup
- Base model: OpenMOSS-Team/MOSS-TTS-Nano-100M (227MB)
- Codec: OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano (85MB)
- Dataset: marcosremar2/gemini-dataset-erinome (9995 wavs, 10000 texts, 4929 valid pairs)
- Train config: per_device_bs=8, grad_accum=4, global_bs=32, lr=5e-5 cosine, warmup=5%, 3 epochs (465 steps), bf16
- Train time: 10min on RTX 4090

## Train metrics
- Loss: 5.5 → 4.7 (decrease over 3 epochs)
- Steps: 465
- Samples/sec: 26.47

## Eval (50-prompt holdout)
- Whisper-base ASR, language=pt, fp16
- WER computed via jiwer

| Model | WER | N |
|---|---|---|
| Baseline | 0.5316 | 50 |
| SFT | 0.1537 | 50 |
| **Δ** | **-0.3779 (-71%)** | |

## Files
- Checkpoint: ./checkpoints/iaratts-sft-v1/checkpoint-last
- Eval prompts: /root/iaratts/eval_prompts.jsonl
- Eval audios: /root/iaratts/eval_out/{baseline,sft}/