bitmanagerai/Qwen3TTSv2_SFT2

Qwen3TTSv2 SFT2

Russian Qwen3-TTS SFT2, released as a full merged checkpoint. It starts from bitmanagerai/Qwen3TTSv2_SFT1 and merges the selected agreement >= 0.95 LoRA checkpoint (step 120,775) into the model weights.

This repository contains ordinary Qwen3-TTS weights, not a LoRA adapter. Load it directly with Qwen3TTSModel; do not attach a PEFT adapter.

Validation comparison

The selected checkpoint was evaluated on 2,000 Russian hard-number utterances. All values are percentages on the 0–100 scale (lower is better). SFT2 improves each transcription-fidelity metric relative to SFT1.

Metric	SFT1	SFT2	Change
Utterance CER	1.466153%	0.522076%	-0.944077 pp
Utterance WER	2.971414%	1.709105%	-1.262309 pp
Number CER	0.891371%	0.308877%	-0.582494 pp
Number WER	1.604707%	0.855844%	-0.748863 pp

SFT1 metrics are reproduced from its model card. SFT2 used the fixed 2,000-row evaluation fingerprint 19d5967fa9a56e1a4104b697d9ce027e0ad987d73ad025a18d7103e55c404ead at epoch 2, optimizer step 120,775.

Evaluation accounting	Count
Evaluated utterances	2,000
Generation failures	0
ASR failures	0
Empty hypotheses	0
Other failures	0

Load as a normal Qwen3-TTS model

import torch
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "bitmanagerai/Qwen3TTSv2_SFT2",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    attn_implementation="eager",
)

The release was verified with this standard loading path. It reports model type qwen3_tts and contains no lora_* parameters.

Training

Setting	Value
Initial checkpoint	bitmanagerai/Qwen3TTSv2_SFT1 (merged step 178,080)
Original Qwen base	Qwen/Qwen3-TTS-12Hz-1.7B-Base
Training data	Prepared Russian Qwen tokenizations with agreement >= 0.95
Adaptation	LoRA rank 16, alpha 32; merged for this release
Learning rate	1e-5
Effective batch	42 (7 GPUs × batch 3 × accumulation 2)
Training length	3 epochs; 1,848,818 prepared rows; 131,754 optimizer updates
Released checkpoint	Epoch 2, step 120,775
Adapter SHA-256	f9c6bad816dd7fbc86874816cd70618c0a24db1f0d065f6d0698c06de18038fc

Notes

This is a text-to-speech model. Quality can vary by prompt, reference audio, and deployment settings. The reported hard-number metrics measure Russian transcription fidelity; they are not a complete measure of speaker similarity, prosody, safety, or general synthesis quality.