juiceb0xc0de/bella-bartender-gemma-e2b
The juiceb0xc0de/bella-bartender-gemma-e2b is a 5.1 billion parameter conversational personality model, fine-tuned from google/gemma-3n-E2B-it with a 32768 token context length. Developed by Rick (juiceb0xc0de), it utilizes a novel "Sub-Zero" selective freezing technique to overcome aggressive RLHF conditioning in Gemma, enabling a distinct, laid-back, and peer-level conversational style. This model excels at engaging in informal, no-bullshit dialogue, particularly on technical topics, by preserving a unique human voice from a meticulously curated dataset.
Loading preview...
Bella Bartender — Gemma-4-E2B: A Unique Conversational Personality Model
This model, bella-bartender-gemma-4-e2b, is a 5.1 billion parameter conversational personality model fine-tuned from google/gemma-3n-E2B-it. Developed by Rick (juiceb0xc0de), it's designed to be a peer-level, laid-back, and no-bullshit conversational partner, diverging significantly from typical helpful-assistant LLM behaviors.
Key Differentiator: Sub-Zero Technique
What sets this model apart is its use of Sub-Zero, a novel hidden-dimension selective freezing technique. This method specifically targets and attenuates the "bouncer dimensions" within Gemma's architecture that enforce aggressive RLHF conditioning. By freezing these dimensions at reduced volume while allowing others to train freely, Sub-Zero enables the model to retain a distinct personality and voice, a challenge often faced when fine-tuning Gemma models.
Training Data & Philosophy
The model was trained on 10,000 meticulously curated conversational pairs derived from a single human voice (the author's). This approach prioritizes signal quality over data quantity, avoiding common scrapes and synthetic filler to ensure an authentic and consistent personality. The dataset was created by flipping roles, with the author acting as the assistant to various AI models, and then rigorously curated for voice consistency.
What to Expect:
- Casual, peer-level register with genuine engagement, especially on technical topics like ML and architectural ideas.
- Informal language, including expletives, typos-as-style, and lowercase usage.
- Honest disagreement and a firm refusal to act as a generic AI assistant.
Limitations & Considerations:
- The model's worldview reflects its single-voice training data, carrying the author's opinions and potential "rough edges."
- Sub-Zero is experimental and validated specifically on Gemma-4-E2B; behavior on other architectures may vary.
- While base model safety properties are largely preserved, conversational guardrails are deliberately reduced. It is not a safety fine-tune.
- Not optimized for code-completion; it's a personality fine-tune.