shiningstar1128/llm-trainer-v09

Qwen3 voice-design TTS for the Vocence subnet.

Hugging Face repo id

model_name in vocence_config.yaml, on-chain commit, and chute wrapper VOCENCE_REPO must all be:

shiningstar1128/llm-trainer-v09

Validator /speak contract

Validators POST JSON (natural-language prompts after source extraction):

{
  "text": "<transcription to synthesize>",
  "instruction": "An adult female with an American accent, speaking at a normal pace in a mid-range pitch, sounding neutral and formal throughout"
}

• text — passed verbatim to generate_voice_design(text=...) (length cap only).
• instruction — passed verbatim to generate_voice_design(instruct=...) (length cap only).

No rewriting, paraphrasing, or SSML enrichment of either field.

Repo layout

File	Purpose
miner.py	Qwen3 TTS engine (Miner class)
vocence_config.yaml	model_name + runtime/generation limits
chute_config.yml	Chutes image, GPU (24 GB VRAM), qwen-tts
config.json	Model config
model.safetensors	Main weights
speech_tokenizer/model.safetensors	Speech tokenizer weights
tokenizer.json, tokenizer_config.json, vocab.json, …	Text tokenizer

Deploy

1. Upload this folder to shiningstar1128/llm-trainer-v09 and pin a 40-char commit SHA.
2. Build/deploy: deploy/deploy_chutes_09.py (VOCENCE_REPO must match this repo).
3. Commit: uv run vocence miner commit --model-name shiningstar1128/llm-trainer-v09 --model-revision <sha> --chute-id <uuid>.

See miner_sample/MINER_GUIDE.md for full miner rules.