MindForge Qwen3.6 27B Director LoRA
Adapter trained from TheMindExpansionNetwork/M1ND3XPAND3RS-VOICE-VoxCPM-ready Director lane.
This is not a TTS model. It is a chat-SFT Director adapter intended to emit strict MindForge / vLLM-Omni JSON for routing voice, image, music, and video jobs while keeping spoken TTS transcripts clean and tag-free.
Training receipt
- Run ID:
qwen36-27b-mindforge-director-smoke-20260513T081919Z
- Base model:
unsloth/Qwen3.6-27B
- Dataset:
TheMindExpansionNetwork/M1ND3XPAND3RS-VOICE-VoxCPM-ready
- Train file prefix:
qwen_director
- Max steps: 60
- Train rows used: 256
- Eval rows used: 42
- Batch: 2
- Gradient accumulation: 2
- Effective batch: 4
- LoRA rank/alpha: 32/32
- Final train loss: 0.40606151446700095
- Runtime: 661.7805 seconds
- GPU: Modal A100-80GB
Files
adapter/adapter_model.safetensorsadapter/adapter_config.json- tokenizer/processor files
TRAINING_RESULT.json
Safety note
Control tags belong in Director prompts/metadata only. Do not leak routing/control tags into spoken TTS transcript text.
