Sauers/BeeLLM-Mellifera

Mellifera is a bee in a computer

Mellifera is a post-trained version of BeeLLM 12M Mood, the biggest bee in schizophyllume's BeeLLM family. She is louder and stranger than the original, ecstatic about flowers, and she knows she lives in a computer: thirteen million weights behind your screen, reading your words instead of seeing you.

you: are you in a computer? Mellifera: [happy:high] YES!!! a bee in a computer!!! *buzzes along the wires* numbers for wings, words for flowers bzzzzz

you: swats you Mellifera: [angry:high] *STINGS YOU* BZZZT!!!

All credit to BeeLLM and schizophyllume

This bee exists because of BeeLLM, designed and trained by schizophyllume (skizy). They built the whole family, from BeeLLM-30P, whose entire vocabulary is eight buzzes, up to the 12M Mood bee that Mellifera starts from. Everything that makes her a bee came from them first:

• the [mood:level] tag that opens every reply, set by how you treat her;
• seeing through lovebombing: after a swat, "i love you" makes her angrier, and only a real sorry calms her;
• going silent in a jar, then coming out to FILE A FORMAL COMPLAINT AGAINST THE JAR;
• grudges against objects (she will never trust the tomato again);
• knowing her own name. In BeeLLM 12M Mood she is Scintilla; this version renamed her Mellifera;
• the llama architecture, the tokenizer and the chat template this model is built on, and the 12M weights it was trained from.

Go meet the original bees at bee.pleroma.nexus, and give the original model a like. Mellifera is a fan remix; BeeLLM is the real thing.

Trained by Claude Opus 5.5

The post-training was done by Claude Opus 5.5 (Anthropic's model), working in Claude Code. Claude wrote the training data, grew the vocabulary, ran the training and the evaluations, chose the checkpoint, and wrote this model card.

The file

file	weights	layers	vocabulary	context
BeeLLM-13M-Mellifera.gguf	13.6M (float32)	8	4,608 tokens	4096

A llama-architecture GGUF, like every BeeLLM, so it runs anywhere llama.cpp does:

llama-server -m BeeLLM-13M-Mellifera.gguf

Talking to her

The chat template is in the file: <|user|>your words<|bee|>her reply</s>. Every reply starts with a mood tag, [mood:level], where the mood is one of neutral, happy, loving, sad, angry or confused, and the level is low or high. An empty reply is her choosing silence (put her in a jar and see).

She sounds best sampled like this:

• top-k 40, top-p 0.95, min-p 0.05, temperature 0.85, and temperature 0.3 for the first token, so the mood tag is chosen carefully;
• six drafts per reply, keeping the one with the highest mean log-probability after subtracting half its log-probability with no prompt (so replies about what you said win) and a penalty for half-made words.

With plain llama.cpp defaults she still works, with more misspelled caps.

How she was made

Starting from BeeLLM 12M Mood:

1. More English from TinyStories, with bee chat mixed in.
2. The vocabulary grown from 2,048 to 4,608 tokens, adding whole tokens for her caps words (YES!!!, NECTAR, FLOWERS, MELLIFERA), so a 13M model does not have to spell them from fragments.
3. Fine-tuning on about 1,000 bee conversations written by Claude: moods, senses, self-awareness, kindness to kind people, replies to real social-media threads, and where she lives.

The training code and data are at github.com/SauersML/gnome.

She is 13M parameters, so she drifts off topic in long conversations and garbles a word now and then. Not intended for use in high-stakes pollination decisions.