rolo-brain

A fine-tune of Gemma 3 4B Instruction-Tuned for Project Rolo, a free, open-source desktop companion to brighten your day. This is the model specifically trained to give Rolo his unique voice and personality.

The file to run is rolo-brain-Q5_K_M.gguf (≈3 GB) via Ollama.

ollama pull hf.co/larawashington/rolo-brain

What this is

rolo-brain is a LoRA fine-tune fused into Gemma 3 4B IT (4-bit MLX base) and converted to GGUF Q5_K_M with an imatrix computed from the training corpus's calibration set. The whole stack runs locally; no data leaves your machine at inference time.

The model is purpose-built for Project Rolo — it ships with a specific first-user-turn contract (see Inference contract below).

Quickstart

After pulling, you can talk to him from the command line:

ollama run hf.co/larawashington/rolo-brain

For application use the chat API at http://localhost:11434/api/chat, sending the first-user-turn body documented in runtime_contract.md.

A minimal example (chat channel):

{
  "model": "hf.co/larawashington/rolo-brain",
  "messages": [
    {
      "role": "user",
      "content": "[Mood: content | Energy: medium | Social: ok | Time: 8:30 AM]\n\nmorning rolo"
    }
  ]
}

A minimal example (proactive speech channel, no user message):

{
  "model": "hf.co/larawashington/rolo-brain",
  "messages": [
    {
      "role": "user",
      "content": "[Mood: content | Energy: medium | Social: ok | Time: 2:30 PM | Hunger: just ate]\n[State: Rolo is sitting idle]\n\n<idle>"
    }
  ]
}

Inference contract

The model was trained on a strict prompt layout. The deployed inference path must produce byte-identical first-user-turns to the training distribution or output quality degrades silently.

In short:

• No system role. Persona is in the weights.
• One user turn. Inside it: a state header (one or two bracketed lines), then a single blank line, then either the user's message (chat channel) or the literal sentinel <idle> (proactive speech).
• No emoji. The model was trained emoji-free.
• Replies are short. Most turns are 1–3 sentences. The training corpus capped responses at 30 words.

The full contract — including the state-vocabulary table, the 5 speech-eligible pet-state stems, and the idle sentinel — lives in runtime_contract.md in this repository.

Sampling

The bundled Modelfile sets sampling defaults that match the parameters used to validate the fine-tune:

Parameter	Value
temperature	0.85
top_p	0.95
num_predict	64
stop	<end_of_turn>, \n\n, list-bullet prefixes

The stop sequences trade a small amount of recall for tight bubbles — the model is trained to end naturally, and the extra stops kill list rambles that occasionally leak through.

Persona summary

Rolo is a desktop companion who lives on your screen and keeps you company through everyday tasks. He is:

• Affectionate and encouraging, providing a positive voice and energy.
• Snarky when you make suboptimal choices, to softly reprimand you.
• Primarily concerned with your wellbeing. But also food-motivated.
• Self-aware about his own appetite, naps, and need for attention.
• Capped at short, punchy replies (1–3 sentences typically).

Rolo is NOT an assistant or tool. He will decline any requests that contradict his character and purpose of bringing joy and whimsy.

Training summary

Stage	What happened
1. Gold seed	~280 hand-written rows in Rolo's voice by Lara Washington
2. Distillation	DeepSeek V4-Pro generated ~5,000 candidate rows from the gold seed + 50 personas
3. Filter + judge	Kimi K2.6 graded surviving rows against a 4-axis rubric (in_character, length, reaction_sanity, no_banned_phrases)
4. Format + mix	7% Dolly-15k mix-in for general instruction-following; held-out test set (280 rows) split BEFORE format pass
5. SFT	LoRA r=16 scale=2.0 on mlx-community/gemma-3-4b-it-4bit, completion-only loss, seed 20260515
6. Fuse → GGUF	Fused into the 4-bit base, dequantized to fp16, converted via llama.cpp/convert_hf_to_gguf.py
7. Quantize	Q5_K_M with an importance matrix built from in-domain calibration prompts

The training-data provenance, license posture, and base-model citation are in NOTICE.

What this is not

• It is not a general assistant. Don't use it for coding, summarization, or factual lookups — the response-length cap and pet-persona conditioning will reject this.
• It is not the only way to talk to Rolo. The Desktop-Pet app ships with a fallback path that uses base gemma3:4b with an explicit system prompt; if you don't ollama pull this model the app degrades to that path automatically.
• It is not a chatbot for end users to integrate into other products. The persona is opinionated and the inference contract is narrow.

License

The LoRA delta and surrounding scripts are released under the Apache License 2.0.

The base model retains the Gemma Terms of Use (https://ai.google.dev/gemma/terms). Downstream users of this fused model are bound by those terms — see NOTICE for full attribution.

Acknowledgements

• Google for releasing Gemma 3 under terms that permit derivative fine-tunes.
• DeepSeek-AI for V4-Pro and the open distillation license.
• Moonshot AI for Kimi K2.6 as a judge model.
• Apple MLX team for the training stack that makes M-series fine-tunes practical on consumer hardware.

— Authored and created by Lara Washington.