e88 88e d8
d888 888b 8888 8888 ,"Y88b 888 8e d88
C8888 8888D 8888 8888 "8" 888 888 88b d88888
Y888 888P Y888 888P ,ee 888 888 888 888
"88 88" "88 88" "88 888 888 888 888
b
8b,
e88'Y88 d8 888
d888 'Y ,"Y88b 888,8, d88 ,e e, 888
C8888 "8" 888 888 " d88888 d88 88b 888
Y888 ,d ,ee 888 888 888 888 , 888
"88,d88 "88 888 888 888 "YeeP" 888
PROUDLY PRESENTS
SorcererLM-8x22b-iMat-GGUF
Quantized with love from fp16 using the alpha=32 version.
Original model author: rAIfle
- Importance Matrix calculated using groups_merged.txt in 105 chunks, n_ctx=512, and fp16 precision weights
Original model README here and below:
SorcererLM-8x22b-bf16
Oh boy, here we go. Low-rank (r=16, alpha=32) 16bit-LoRA on top of WizardLM-2-8x22B, trained on 2 epochs of (cleaned & deduped) c2-logs. As far as I can tell, this is an upgrade from WizardLM-2-8x22B for RP purposes.
Alongside this ready-to-use release I'm also releasing the LoRA itself as well as the earlier epoch1-checkpoint of the LoRA.
Why A LoRA?
The choice was fully intentional. I briefly considered a FFT but for this particular use-case a LoRA seemed a better fit. WizardLM-2-8x22B is smart by itself but its used vocabulary leaves much to be desired when it comes to RP. By training a low-rank LoRA on top of it to teach it some of Claude's writing style, we remedy that.
Prompting
- Use the templates in Quant-Cartel/Recommended-Settings under the
SorcererLM-folder. - Or Vicuna 1.1 and a sane context template. It's somewhat sensitive to samplers, I'd recommend Temperature 1, MinP 0.05 and a dash of DRY but YMMV. Shorter prompts seem to work better, too.
Quantized Versions
Acknowledgments
The main shoutout I want to make is to my Cartel bros, Envoid and particularly I^2, for being amazing. I count this as a team effort, so they deserve kudos too if you like this.
Training
Trained using qlora-pipe. Configs included in the train-subfolder.
Safety
... n/a