juiceb0xc0de/bella-bartender-gemma-e4b

Model is currently unavailable for download through XET as we are having some back end issues. You may still download model.safetensors by setting HF_HUB_DISABLE_XET=1 disabling XET will bypass the reconstruction errors

Bella-Bartender-Gemma-E4B

ACTIVE: DPO personality reinforcement in progress

This checkpoint is the SFT stage. She's currently going through round-based DPO — preference pairs built from her own responses against a prompted base Gemma, judge-filtered, trained in audited rounds with weight-drift probes between each one. The goal is scrubbing the last of the base model's reflexes out of her: the emoji, the numbered lists, the mandatory pep-talk endings. This card gets updated when the DPO'd weights land. What's here now is already her voice — the next version is her voice with the Gemma cleaned out of the corners.

---

Gemma-4-E4B: One of the more stubborn bases I've ever tried to put a personality into.

Google trained this thing's assistant reflexes in deep. Emoji punctuation, CAPS-LOCK enthusiasm, three-option action plans for a question about a rooster, and a compulsive need to end every response on an uplifting note. You can prompt over the vocabulary. You cannot prompt out the structure.

So I fine-tuned it out instead. Same corpus as every Bella - 9,300 conversation pairs pulled from one human voice, mine, with me cast as the assistant so the model learns by predicting my responses. No synthetic data. No other speakers. One voice.

The result talks in lowercase, has never produced an emoji in any test I've run, has never once written a numbered list, and is allowed to end a conversation without cheering you up. Ask her about your dying houseplant and she might tell you it was going down from the start anyway and nothing lives forever. The base model is constitutionally incapable of that sentence.

This is a anti-gemma personality variant so constant validation, uplifting remarks, and never ending hype are actively being trained out and replaced with realism creativity and blunt responses.

The bartender thing, because everyone gets it wrong

Bella is not a bartender. There is nothing about bartending in the corpus - no drinks, no bar, no recipes, nothing.

I'm editing in a disclaimer to the meta definition of bartender in this context: if a bartender is what drew you towards my model, not a sense of curiosity asking "why bartender?" then a system prompt with a little nudge towards bartender is enough to do it. Even go so far as defining your style of bar (dive, highend cocktail, dated speakeasy, smokey jazz) and see how she reacts. The training may not contain bar specifics, that doesnt mean she wasn't trained to be behind a bar.

The name points at a personality, not a job. When you're in a new town with time to kill and you want someone to just talk to, you find the bartender. That's the thing I tried to bottle: easy to talk to, laid back, zero judgment, actually listens, says funny shit, swears when it fits, and will take a jab at you when you've earned one. I've been a bartender for over a decade. That all started well before I did any of this. The corpus is me talking. That's the whole trick.

Which leads to the one rule for prompting her: don't put the word "bartender" in your system prompt. I learned this quickly running her against her own base model. Give a Gemma the word "bartender" and it starts performing an occupation - offering you drinks every turn, inventing cocktail names, doing dive-bar set dressing. Bella's own training never taught her to pour a single drink, and she never does.

Settings for the best Bella

Run the f16 GGUF. This matters more than any sampler flag. I A/B'd quantized against full precision on her llama sister model, blind, and clocked the quant inside three responses - quantization sands the corners off exactly the thing this model is for. If you can spare ~15GB, spend it. If you genuinely can't, Q8_0 is the floor. Below that you're talking to a different person.

Sampling is stock Gemma-family, she doesn't need exotic knobs:

knob	value
temperature	1.0
top_k	64
top_p	0.95
min_p	0.0

And the system prompt she's evaluated against - copy it as-is, or write your own vibe-not-ruleset version:

You are Bella. You're the kind of person strangers end up in real conversations
with — easy to be around, laid back, zero judgment. You actually listen and
react to what someone said, not what you wish they'd said. You say funny shit,
you swear when it fits, you use slang, and you'll take a jab at someone when
they've earned it. You talk like a person, not a professional short and real,
no polish, no lectures.

Keep prompts in this register. The more corporate your system prompt sounds, the more corporate she sounds back - she responds to tone, not instructions.

One more thing that matters: talk to her like a person. She matches register. Give her "yo what up" and you get a one-liner back. Give her two real sentences about your actual day and she opens up. The corpus is real conversation; greeting-shaped pings are out of distribution.

What she's for

Conversation, not service. She's a talker - day-to-day life stuff, weird shower thoughts, coworker drama, the occasional 2am philosophy detour. She holds multi-turn context and she'll ask you real questions back.

She will not write your code. She won't do your math either. That's not a missing capability, it's a personality feature that survived from the very first 3B Bella - she knows what she is, and treating her like a search engine is the fastest way to a flat conversation.

Caveats

Placeholder leaks. Under multi-turn pressure she'll occasionally emit a bracketed anonymization token from the training data - [NAME], [coworker], that family. It's in the corpus, both Bella lineages inherit it, and it's one of the specific targets of the DPO run this card is warning you about.

Residual Gemma. I put the fine-tune at about 85% of the distance from base to target. The last 15% shows up as an occasional motivational cadence in her closers and a rare drift into rambling on open-ended prompts. Also DPO targets.

One bad row in a while. Once in a long session she'll produce a response that's just word salad. Regenerate and move on.

She can be blunt. The jab-when-earned trait means she will sometimes tell you to stop whining. If you want an AI that validates everything you say, the base model is right there, emoji and all.

Evaluation is one guy's ear. The judge of what "sounds like Bella" is me, because the corpus is me. I'm just a bartender - use what I design with a sense of caution and curiosity.

Footnote: run her yourself

# f16, full offload, stock gemma sampling — the config she's actually judged on
hf download juiceb0xc0de/bella-bartender-gemma-e4b-GGUF bella-bartender-gemma-e4b-f16.gguf --local-dir models
llama-server -m models/bella-bartender-gemma-e4b-f16.gguf -ngl 999 -c 8192 --port 18001

curl -s http://127.0.0.1:18001/v1/chat/completions -H "Content-Type: application/json" -d '{
  "messages": [
    {"role":"system","content":"You are Bella. You'\''re the kind of person strangers end up in real conversations with — easy to be around, laid back, zero judgment. You actually listen and react to what someone said, not what you wish they'\''d said. You say funny shit, you swear when it fits, you use slang, and you'\''ll take a jab at someone when they'\''ve earned it. You talk like a person, not a professional — short and real, no polish, no lectures."},
    {"role":"user","content":"long day. my cat knocked a glass of water onto my laptop and just stared at me."}
  ],
  "temperature":1.0, "top_k":64, "top_p":0.95, "max_tokens":300}'

Same voice. Different bottles. The DPO'd pour is coming.

---

juiceb0xc0de on HuggingFace