outstanding first impression
I just like to let you know that my first impression of this blend of large language model is outstanding
It seems to understand the complex relationships of what I write to it with astonishing accuracy, and its answers are precise and straight to the point. It can react somewhat reservedly to simple queries, which is beneficial for its efficiency and clarity; it then easily build on what has been written so far in the dialogue and become even clearer if needed. It feels like it is even able understand complex topics such as Kant's ethics and the fatal social implications when, for example, a hypothetical imperative is mistakenly passed off as a categorical imperative with so much depth that it can provide seemingly infinite material for writing lively and thought promoting satires.
I used only the Q8_0 version, maybe even more so I made a mistake with other models I tried when I sought for largest possible models but therefore used a lower Q version to fit the available hardware, but even so, when I first tested it, it felt like this model here was tailor-made for me.
Thanks a lot for letting us now! Let's hope @DavidAU sees this :) I'm curious to try it out now :)
Thank you
@luminonaut
@mradermacher
;
This model mixture (and others like it) are designed to address a few core issues:
- Fine tunes are excellent, but losses from tuning impact overall function and some loss of knowledge.
- Instruct models are excellent at following instructions, but lack fundamental knowledge... and character.
- Layers ARE function, depth, nuance and knowledge. The more you have to more you can do.
By bringing the instruct back (which improves both function, and brings back most knowledge lost during tuning) and the fine tuned model (75% of it) well.. it is potent combo.
There are now several instruct/fine tune models at my repo (Llama3 "L3", Llama3.1, Mistral Nemo) and more are coming.
The "Brainstorm" models (GGufs at my repo, source code uploading shortly) take these principles even further; brainstorm is all about forcing the model to "ponder"
and "reconsider" - but at the root core level.
REPO: https://huggingface.co/DavidAU
I am a little... ahh.. behind in uploading source code right now; the upload gods are fickle here.
RE: Blackroot; interestingly (based on your comments) this model is one of the core models in my "horror" series.
It is an exceptional fine tune.
Thank you again for the feedback ;
Thanks for the interesting background details.
Whow, "horror series", at first a fun fact but explains a lot to me when i think about it.
With common models i often needed to explicitly prompt it to not wash out drama from their responses when trying to generate a frankly text about something that might be not matching an exaggeratedly "happy world". I assume the extemes of horror gives a constructive spin that allows the blackroot models to respond straight on without drifting into such unrealistic exaggeration. And gives an unexpected perspective.
That might also be beneficial to mitigate confirmation bias and other biases which are highly likely present in all larger datasets.
100% ; the "normal" models have a "sunshine and puppies" / "positivity" bias - which many times is so strong it obscures reality, never mind the "facts".
This really shows when you want something uncensored like brainstorming processes and/or any creative process.
A lot of "fine tunes" have a negative bias to counterbalance this issue.
Some... go overboard.