Phase 1b re-run: J-Moshi-ext + Path B synthetic dialog FT

Fine-tuned from nu-dialogue/j-moshi-ext with 100 synthetic dialogs (Irodori-TTS Full FT amitaro voice, generated 2026-07-04) for voice-transfer validation in the Path B research programme.

Training config

Base model: nu-dialogue/j-moshi-ext (with --extend_modules_for_user_stream → dep_q=16 during train)
Data: 100 synthetic dialogs (612 turns, 31.7 min stereo 24kHz)
Params: all 7.5B trainable (PARAMS_TO_FT=all)
Steps: 65 (5 epoch × 13 iter, effective batch 8)
LR: 3e-5 (tempformer + depformer)
DeepSpeed: ZeRO-3 + bf16 + CPU offload optimizer
GPU: A100 SXM4 40GB × 2 (Slovenia vast.ai, $1.334/hr)
Wall clock: ~1 h train + ~7 min consolidate
Final loss: 1.27 (text 0.48 + audio 0.79)
Cleaned: dep_q 16 → 8 via tools.clean_moshi --remove_modules_for_user_stream

Files

model.safetensors (15 GB, bf16, dep_q=8) — inference-ready
moshi_lm_kwargs.json — moshi LM config

Downloads last month: -

Safetensors

Model size

8B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ayousanz/moshi-phase1b-rerun-2026-07-06

Base model

kyutai/moshiko-pytorch-bf16

Finetuned

nu-dialogue/j-moshi-ext

Finetuned

(4)

this model