juiceb0xc0de/bella-bartender-gemma-e2b

Text Generation • 5B • Updated May 11 • 8 • 4

Note The latest model in the Bella Bartender lineup, trained using Sub-Zero, a method I'm developing for dimension-level weight surgery. Sub-Zero triangulates weight matrices that encode corporate assistant language patterns. I take phrases like "As an AI model," "No fluff," or "Why this matters" and freezes those specific directions with gradient hooks before fine-tuning. Everything else trains normally. The result is a Bella that doesn't have the RLHF voice baked underneath it's surgically removed.

juiceb0xc0de/bella-bartender-8b-llama3.1

Text Generation • 8B • Updated 13 days ago • 1.66k • • 4

Note 🍹 Bella Bartender v1 🍹 model who is laid back, gets along with everyone. She was trained on a single voice, so her grasp of metaphor and complex ideology is uniquely shaped by one perspective. See how far you can push her opinions before she pushes back. Importance Matrix (I-Quants) (IQ2_XS to IQ4_XS) and Standard K-Quants (Q3_K_M to Q8_0) models are available.

juiceb0xc0de/bella-bartender-heretic-1b

Text Generation • 1B • Updated 13 days ago • 2.5k • 2

Note The heretic versions of Bella Bartender were done post SFT to preserve the models adherence to the single voice. Training on models that are pre abliterated proved themselves to be less creative or authentic as the SFT > abliteration models. Changing the workflow saved the idea of having abliterated models in the Bella family. My findings were that allowing the model to learn with guidance kept the voice tighter and more natural. Post SFT abliteration had littler effect on quality of character.

juiceb0xc0de/bella-bartender-3b

Text Generation • 3B • Updated 13 days ago • 1.34k • 2

Note 🍺 Bella Bartender 3B 🍺 The original bartender model punches far above her weight class on complex ideas. She will make you question why you ever needed more parameters. Trained on a single voice rather than a corpus of scraped data demonstrates how much perspective matters in LLM training especially at low parameter counts. Importance Matrix (I-Quants) (IQ2_XS to IQ4_XS) and Standard K-Quants (Q3_K_M to Q8_0) models are available.

juiceb0xc0de/bella-bartender-v2-8b

Text Generation • 8B • Updated 13 days ago • 670 • 4

Note 🍷 Bella Bartender v2 🍷 after Bella Bartender v1 surpassed all expectations with Sonnet 4.6 LLM as a judge benchmarks I decided to take all the critiques and see if I was able to pull datasets from my collection that would make her even more versatile and life-like. V2 has the same attitude and punch as the first however she plays more like the 2 AM version. The Bella who's still awake with you at the end of the night but doesn't have time for your bullshit.