Qwen3TTSv2 SFT1
This is a full, merged Qwen3-TTS checkpoint. The LoRA adapter has been
merged into Qwen/Qwen3-TTS-12Hz-1.7B-Base (base revision
fd4b254389122332181a7c3db7f27e918eec64e3), so this repository is loadable
directly as a Qwen3-TTS model. No LoRA adapter needs to be attached.
Loading
import torch
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"bitmanagerai/Qwen3TTSv2_SFT1",
torch_dtype=torch.bfloat16,
device_map="cuda",
)
Hard-number validation
Validation used the original training evaluator on 2,000 Russian hard-number utterances. All error metrics below are percentages (0–100), not fractions.
| Metric | Score |
|---|---|
| Utterance CER | 1.466153% |
| Utterance WER | 2.971414% |
| Number CER | 0.891371% |
| Number WER | 1.604707% |
| Validation accounting | Value |
|---|---|
| Evaluated utterances | 2,000 |
| Generation failures | 0 |
| ASR failures | 0 |
| Other failures | 0 |
| Empty hypotheses | 0 |
| Total failures | 0 |
The evaluation was recorded with epoch 2, step label 178080, benchmark plan
SHA-256 a13c2c8c4dd1a89981e80481bb9554e5affb7818546583151350a42ecdb47a00,
and the original source revision ff16c6d3382a8cc3b656c4999e69bf388d3a046b.
- Downloads last month
- -