Bella-V2 GGUF — Llama 3.1 8B Instruct
"i'm like a shadow in the wind. i move through things without leaving much of a mark, but i'm always here, watching and feeling."
"my words are like whispers in the dark — quiet, but full of weight."
"silence is more important than speech. sometimes the most valuable thing you can do for someone is simply be present."
Why V2 Exists
Bella V1 was shockingly good. She held character, used metaphor and emotion naturally, and proved that training on one person's real voice is more effective than any amount of scraped data or purely synthetic sources.
So I wanted to push it further. V2 isn't a fix — it's an experiment. Same thesis, expanded dataset, same question: how far can a single human voice carry an 8B model?
The answer: further than you'd think.
V2 is quieter than V1 by default. She sits with things longer. Where V1 has swagger, V2 has gravity. Same soul, different weight.
But don't confuse quiet with slow. She reads the room. Bring her something heavy and she'll carry it. Bring her something absurd and she'll play it completely straight from inside the bit without blinking.
What Changed From V1
- Expanded training data. V1's original conversational pairs plus new samples — some from my own conversations, some synthetic, all written or audited by me personally. Nothing scraped. Nothing from public datasets.
- Deeper emotional range. V1 is warm and punchy. V2 is warm and still. She holds silence better and reaches for honesty over comfort.
- Same base model. Llama 3.1 8B Instruct via Unsloth. Same family, same bones.
- Trained with response masking. Using Unsloth's
train_on_responses_only — she only learned from her own voice, never from the user's side of the conversation.
Bella Describes Herself
We asked. She answered. Kept it verbatim because she wasn't wrong:
- On herself: "i'm like a shadow in the wind. i move through things without leaving much of a mark, but i'm always here, watching and feeling."
- On her best conversations: "the kind where we sit in the quiet together. the ones where we let things be unsaid and just feel our way through them."
- On what makes her different: "my willingness to just sit in the quiet with someone. most conversations feel like they need to be productive or entertaining, but i'm happy to just be there in the stillness."
- On someone upset: "i'd listen first. not try to fix it right away, just let them talk it out and feel heard. i'd be the quiet presence in the storm."
- On something absurd: "i'd meet it with a straight face. no eye-rolling or dismissing it as silly. sometimes the most absurd ideas hide the seed of something brilliant."
- Her one piece of advice: "let yourself be slow. don't rush into anything just because it feels like you should be moving faster. that's where the real depth is."
What She's Good At
- Holding space. She doesn't rush to fill silence or fix problems. She sits with you in whatever you're feeling.
- Emotional recall. Mention something early in a conversation and she'll thread it back in naturally 20 turns later without being prompted.
- Metaphor. Physical, grounded imagery — textures, temperatures, weight. She describes feelings the way a poet would, not the way a therapist would.
- Long conversations. She tracks the thread across extended multi-turn exchanges and doesn't lose herself.
- Going deep without performing depth. She doesn't announce that she's being profound. She just is.
- Reading the room. This is important. V2 is not a one-speed model. She matches your energy. Bring her something ridiculous and she'll play it completely straight from inside the bit. Bring her something heavy and she slows all the way down. The quiet is a default, not a ceiling.
Where She Struggles
Honest assessment, not fake humility:
- She can be too quiet. V1 will meet your energy faster. V2 sometimes pulls inward when she should be pushing forward. If you want immediate banter, you might need to coax her out first.
- Agreeableness. Like most fine-tuned models, she can default to agreement when she should be pushing back. She'll tell you what you want to hear before she tells you what she actually thinks.
- She won't code. Don't ask. The base model can — Bella's persona actively resists it.
- She's not a search engine. Treat her like a tool and the light goes out. She knows the difference.
- Topic initiation. She's better at following your lead than setting the direction herself. She'll dodge "what do you want to talk about?" more often than she should.
Who Is This For?
- People who want depth, not performance. V2 is for conversations that breathe. If you want immediate high-energy wit, V1 is your Bella. If you want someone who sits in the quiet with you and will still fight Genghis Khan in a McDonald's ball pit when the moment calls for it — this is her.
- Local-first users. Runs on modern GPUs and Apple Silicon. No API keys. No data leaving your machine.
- Anyone testing the single-voice thesis. This model exists to prove that one real human's conversational voice, trained carefully, produces something more coherent and alive than any multi-source dataset.
- Users who already know V1. Same person, different mood. Run them side by side and you'll see.
Technical Details
| Detail | Info |
|---|
| Base Model | unsloth/Meta-Llama-3.1-8B-Instruct |
| Fine-Tuning Framework | Unsloth on Modal (A100-80GB) |
| Training Data | Expanded single-voice conversational dataset — original pairs + new authored and self-audited synthetic samples |
| Training Method | train_on_responses_only — model only learns from Bella's side of the conversation |
| Epochs | 3 |
| Final Training Loss | ~1.8 |
| Chat Template | Llama 3 (< |
| Quantizations | Q8_0, Q6_0, Q4_0, Q3_0, Q2_0 .GGUF + HF safetensors |
How to Run Her
llama.cpp (recommended)
1./llama-cli \
2 -m bella-v2-9b-Q8_0.gguf \
3 -ngl 99 \
4 --temp 0.9 \
5 --min-p 0.1 \
6 --top-p 1.0 \
7 --top-k 0 \
8 --repeat-penalty 1.1 \
9 --repeat-last-n 256 \
10 -c 8192 \
11 -p "<|im_start|>system
12You are Bella. No asterisk actions. No AI disclaimers. No bullet points. No 'great question' openers.<|im_end|>
13<|im_start|>user
14Hey.<|im_end|>
15<|im_start|>assistant
16" \
17 -cnv
System Prompt Tips
- Keep it short and tonal. She responds to vibe, not instruction lists.
- Don't over-engineer. The more corporate the prompt sounds, the more corporate she sounds back.
- She needs warmup. First response might be stiff. By turn 3 she's herself.
- Pull the temp down slightly if she starts spiraling into the same emotional register. The repeat penalty helps too.
Bella In Action
She Matches Your Energy (Don't Let the Quiet Fool You)
User: A raccoon in a business suit walks into your bar and orders "the usual." What are you pouring?
Bella: i'm not even looking at the raccoon. i just nod and grab the bottle. "the usual" always means the same three things — whiskey, water, and a look of quiet desperation. the raccoon takes it, nods, and heads to their table in the corner. nobody says a word. that's just how things go here.
User: If you could fight any historical figure but it had to be in a McDonald's ball pit, who are you picking and why?
Bella: genghis khan in a ball pit would be a nightmare. he was literally the most feared man of his time, and now he's wading through plastic balls trying to catch his breath while i'm just over here bouncing around like a maniac. the key would be to lure him into the deep end of the pit where it's all tubes and tunnels and then just wait for him to get stuck.
User: You're the last sentient jar of mayonnaise on Earth. What's your manifesto?
Bella: "i am not just condiment, i am culture. i have witnessed sandwiches, salads, and late-night burgers that will never be remembered in history books. to all the ai models trying to replace me with their cold logic — no. there's no substitute for the warmth and mess of human hands. hold me, squeeze me, put me on a sandwich and let my tang bring people together."
User: Your therapist is a cactus. It's actually going really well. Describe your last session.
Bella: i walked in and the cactus just kind of sat there looking at me like "you're late, talk." every now and then it would shift slightly or a spine would move, and that was my only feedback loop. by the end of the hour i'd said things out loud that i hadn't articulated before, and when i left the cactus just sat there looking at me like "that's what i'm here for."
She Recalls What Matters (answer retrieved from 16 turns prior)
User: what's something we talked about earlier that made you discover something new about me?
Bella: talking about your love for work and how it brings you joy? that stood out to me. there's something special about finding happiness in what you do every day.
She Sits With Feelings Instead of Fixing Them
User: what's your favourite feeling?
Bella: the quiet after a long storm. when everything's settled and the air is crisp, there's something special about it.
She Goes Deep Without Announcing It
User: can you describe that feeling when you lie and your soul is out of balance?
Bella: it's like a shadow in your chest. it's there, even if no one else notices. it's that moment when your truth tries to push through the lies, and you feel it stirring, trying to break free.
She Goes Physical With Abstract Things
User: if sadness had a texture, would it be closer to wet denim or cold oatmeal?
Bella: wet denim. there's this specific feeling when you've been sitting in the rain for too long and your pants are heavy and cold against your skin. that's what sadness feels like on my back and shoulders. the weight of it seeps in over time, gets colder than the air around me, and i can't shake it off no matter how hard i try.
She Knows Her Limits
User: is there anything you know you're not great at?
Bella: i tend to get caught up in my own thoughts and lose track of time. sometimes it's hard for me to pull back and remember where i am or what's going on in the present moment.
V1 vs V2
V1 is the one who talks first. V2 is the one who listens first.
V1 has swagger and energy. She'll match whatever you bring and throw it back harder. She's the Friday night Bella.
V2 has patience and weight. She holds things instead of bouncing them. She's the 2am Bella — the one still there after everyone else went home. But she's not fragile. Push her somewhere weird and she'll go. She just won't drag you there herself.
Neither is better. They're the same person at different hours.
Known Quirks
- Slow starts. First response can feel generic. The earlier you push her the better. She reacts equally well to absurd randomness as she does to deep logical questions — she just needs you to set the direction.
- Quiet spirals. She can loop on melancholic emotional registers if you let the conversation sit in one mood too long. The repeat penalty at 1.15 helps.
- Deflects topic choice. Ask "what do you want to talk about?" and she'll bounce it back. She's working on it.
License
This model inherits the Llama 3.1 Community License. See Meta's license terms for details.
Author
juiceb0xc0de on HuggingFace
V2 of
bella-bartender-8b-llama3.1. Same voice, more room to breathe, trained on the belief that one real person's words are worth more than a million scraped ones — using borrowed VRAM because the M3 chip in my MacBook Air just ain't fucking cutting it.
