Nov 2, 2024 02:44 Update README.md 02:30 Upload 2 files Very early layerskip-aware weights Oct 26, 2024 01:35 Upload 6 files Added updated LoRAs 01:36 Cleanup mrq
01:27 Update models/config.llama.yaml Reduced a lot of unneeded settings in the config YAML Oct 17, 2024 22:10 Update models/config.retnet.yaml Modified keys 22:09 Update models/config.llama.yaml Modified keys Oct 6, 2024 03:58 Update README.md 03:27 Upload 2 files Updated LoRA (stubborn thing since the base model already handles GLaDOS servicably, but this does help "clean" up things.) 01:44 Upload 2 files Added LoRA (probably the hardest voice for the model, and it's only "okay" before it overfits) 01:39 replaced ar+nar-tts+stt-llama-8 as the base ar+nar-llama-8 since it pretty much surpasses it now mrq
Oct 5, 2024 21:48 Upload fp32.sft Updated weights (post-training to try and clean up the NAR, probably sounds better?) Sep 28, 2024 13:46 Upload fp32.sft Updated weights, again (more post-training, seemed to really need it). Sep 27, 2024 04:00 Upload fp32.sft Yet another updated weights (did some more post training on sampling per group, then culled the dataset to focus more on my own corpus to help bolster it), seems to help? Sep 26, 2024 23:24 Upload fp32.sft Updated weights (received more duration-post-fix training, seems better now) 03:01 Upload fp32.sft Updated weights (the prior ones were "post-fix trained" with an out of date tokenizer) Sep 25, 2024 00:52 Upload 2 files Updated "foundational" model. Sep 8, 2024 23:07 Update README.md 23:00 Upload 8 files 02:18 Update README.md 02:02 Upload 2 files Sep 2, 2024 16:02 Upload fp32.sft Received additional training on my 7900XTX (AMP bfloat16, SDPA):
* a number of unshuffled samples from 12 seconds to 24 seconds, to verify that my ROCm setup works.
* another number of shuffled samples from 3 seconds to 32 seconds, to re-"teach" the model to work for any duration rather than the last duration it was trained against.
* another number of shuffled samples from 3 seconds to 60 seconds but with a RVQ distribution favoring the higher levels, to see if it lobotomizes the AR and clean up the NAR's audio (it seems fine). Aug 19, 2024 04:32 Upload fp32.sft I don't know what the previous weights were but they didn't work, but these do. Aug 10, 2024 00:52 Update README.md RIP the DAC dream Aug 5, 2024 04:19 ugh mrq
Aug 4, 2024 19:50 Update README.md 05:01 Upload 2 files Updating latest copies in safetensor form. Aug 3, 2024 18:20 Update README.md Jul 31, 2024 03:45 Upload 3 files 03:42 stuff mrq
Jul 19, 2024 21:53 Update model/config.llama.yaml 21:48 Upload fp32.pth Updated weights with un-unified position IDs + split classifiers (trained an additional 200k samples each to "fix" them) Jul 1, 2024 19:22 Update README.md actually the 44Khz DAC model seems fine for RVQ levels 0-3 19:11 Upload 2 files 18:44 Upload fp32.pth 18:44 Update README.md Added additional observations from recent experimentation...... Jun 30, 2024 15:39 updated yamls mrq
Jun 16, 2024 12:54 Upload fp32.pth Jun 13, 2024 23:59 Update README.md 23:33 Upload fp32.pth 01:39 cleanup mrq
Jun 9, 2024 22:21 Upload 5 files Uploaded adequate weights 22:15 shuffled things around mrq
Nov 15, 2023 15:15 Upload fp32.pth Oct 27, 2023 18:19 Upload fp32.pth Oct 22, 2023 18:35 Upload fp32.pth Oct 11, 2023 16:13 Upload fp32.pth Oct 10, 2023 22:36 Upload fp32.pth Oct 8, 2023 15:49 Upload fp32.pth Oct 6, 2023 15:05 cleanup mrq
