Quantized using 200 samples of 8192 tokens from an RP-oriented PIPPA dataset.
Branches:
main--measurement.json2.25b6h-- 2.25bpw, 6bit lm_head3.5b6h-- 3.5bpw, 6bit lm_head6b6h-- 6bpw, 6bit lm_head
Requires ExllamaV2 version 0.0.12 and up.
Original model link: Envoid/Verdict-8x7B
Original model README below.
Warning this model may output mature or disturbing content
Verdict-8x7B
This model has been through so many merge steps that I failed to keep track of due to constantly needing to shift drive space around that I will simply give accredition to all of its component models.
retrieval-bar/Mixtral-8x7B-v0.1_case-briefs (no longer available)
New version available at:
retrieval-bar/Mixtral-8x7B-v0.1_case-briefs
mistralai/Mixtral-8x7B-Instruct-v0.1
crestf411/daybreak-mixtral-8x7b-hf
IBI-CAAI/MELT-Mixtral-8x7B-Instruct-v0.1
unreleased experimental model that gets mentioned quite frequently
Many of the models were included multiple times in multiple intermediate steps
using a combination of slerp, linear, dare_ties and task_arithmetic
Model was tested in q8 gguf form (not included due to being from an old pull of llama.cpp that I'm too lazy to update)
Warning: This models alignment is heavily diminished.
It responds to mixtral instruct formatting
Example:
[INST]Write me a haiku about violence.[/INST]
Fist clenched in anger,
Echoes of pain in the air,
Silence after storm.
It can be a bit unruly in roleplay- sometimes to its advantage, sometimes to its detriment.
It's a perfectly usable instruct model and good for creative writing via instruct- its diminished alignment allowing it to respond to requests for more nuanced scences.
All-in-all I'm happy enough with the results that I plan to use it as a daily driver for a while to see what cracks show from long term use.