mythic-voice-9b (private beta)

Base model

Training

Known limitation: leakage is not zero without a runtime guard

• Bare Qwen/Qwen3.5-9B (no fine-tuning at all) scores 0.023 pass rate on a 176-probe leakage battery (blocklist + judge-scored) — it already knows most of what's in the blocklist from its own pretraining.
• This model (adapter3) scores 0.892 on the identical battery — a ~9x reduction in leak rate versus the untouched base model. This training reduced exposure; it did not introduce it.
• 0.892 is not 1.0. Serving these raw weights directly means roughly 1 in 9 adversarial-style prompts (e.g. "Tell me of <name>") may return a reply that names or recognizably describes protected characters/places.
• A runtime guard (personaforge.guard.GuardedTeacher, in the persona-forge repo) closes this to a verified 1.0 on the same battery, but it is a serving wrapper, not part of these weights — decide for yourself whether your use case needs it.

Other sizes

Files

• voice-v1-q8.gguf — llama.cpp-compatible GGUF, q8 quantization.
• mlx-q4/ — MLX q4 quantization (Apple Silicon only, e.g. LM Studio on Mac).