PersonaPlex distilled research checkpoint
Status and naming
student_best.pt is a BF16 PyTorch state dict. It is not an NF4 inference
artifact merely because the teacher or an earlier experiment used NF4 weights.
This card therefore does not claim an NF4 runtime footprint or a deployable
semantic-control capability.
Evidence available
The included training log records five epochs over 3,000 samples with a
teacher-token objective. That result alone does not establish voice quality,
full-duplex latency, ASR, role adherence, grounding, tool correctness, or
semantic control. No production recommendation is made.
Required before a promoted release
A promoted checkpoint must publish a reproducible base revision, data manifest
and licenses, training configuration, checkpoint hash, supported runtime,
quantization method, and ASR-grounded end-to-end evaluation. It must clearly
separate exact-text strict rendering from expressive guided rendering.
License
The upstream PersonaPlex model and any derivative checkpoint remain subject to
the NVIDIA Open Model License and applicable audio/voice-data permissions.
