From ef5792256cf5e04cb9074ced20dbc056039156dc Mon Sep 17 00:00:00 2001 From: Joey Grasty Date: Wed, 2 Sep 2026 17:26:25 -0500 Subject: [PATCH] Add scale-up SFT model Drive links to the README Both Qwen3-30B-A3B-VoxDay-Nuttall-SFT-Scaleup-ep2/ep3 GGUFs are now uploaded to the same Byron_AI_LLMs Drive folder as the original ep2/ep3 checkpoints, for live-testing comparison between the two SFT rounds. Co-Authored-By: Claude Sonnet 5 --- README.md | 22 +++++++++++++++------- 1 file changed, 15 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index 807a7bd..0261d1e 100644 --- a/README.md +++ b/README.md @@ -30,17 +30,25 @@ out. ### 1. Get the model file -Two candidate checkpoints are available. Separability and length-calibration -metrics lean slightly toward epoch 3 (see `JOINT_TRAINING_CLOSEOUT_REPORT.md` -for the numbers), but not decisively enough to settle it on metrics alone — -**live testing through `voice-bench.html` is what should decide between -them.** Download both, run one at a time (or both at once on different -ports — see step 3), and compare generations on real held-out packages -before picking one. +Four candidate checkpoints are available, across two SFT rounds on the same +joint CPT base: the original 208-package SFT, and a newer 433-package +scale-up SFT trained on eight additional untrained books (see the Book +dropdown in `voice-bench.html`). Within each round, separability and +length-calibration metrics lean slightly toward epoch 3, but not decisively +enough to settle it on metrics alone — **live testing through +`voice-bench.html` is what should decide between them**, both within a +round and between the original and scale-up rounds. Download whichever +you're comparing, run one at a time (or several at once on different +ports — see step 3). +Original SFT: - `Qwen3-30B-A3B-VoxDay-Nuttall-SFT-ep3-Q8_0.gguf` (~31GB): https://drive.google.com/open?id=1QZT488Uu1b9Pxoie9Hs2lnxtOhMD4Pwr - `Qwen3-30B-A3B-VoxDay-Nuttall-SFT-ep2-Q8_0.gguf` (~31GB): https://drive.google.com/open?id=1q1w_XMp6oRmfx8xqXDV-HvDx_krVxQtD +Scale-up SFT: +- `Qwen3-30B-A3B-VoxDay-Nuttall-SFT-Scaleup-ep3-Q8_0.gguf` (~31GB): https://drive.google.com/open?id=1nreKTn2JnEtQO2ydrCg8uOJbxcuOIQRO +- `Qwen3-30B-A3B-VoxDay-Nuttall-SFT-Scaleup-ep2-Q8_0.gguf` (~31GB): https://drive.google.com/open?id=1xHeXjLrQyvd_Mey1EvgCz6raeAxJJFzG + Memory needed: ~31GB for the weights plus ~3GB of KV cache at the full 32768-token context, so ~35GB total. Confirmed working on a DGX Spark; should comfortably fit a 48GB L40 as well. If a machine has less than that,