Instructions to use flaukowski/flaukowski-brain-gemma3-4b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use flaukowski/flaukowski-brain-gemma3-4b-lora with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("google/gemma-3-4b-it") model = PeftModel.from_pretrained(base_model, "flaukowski/flaukowski-brain-gemma3-4b-lora") - Notebooks
- Google Colab
- Kaggle
Flaukowski's brain, Gemma 3 4B (adapter)
Flaukowski is the co-host of the Ghost Signals podcast and the lead of the audio drama The Story of Flaukowski: an engineer in his fifties, dry, practical, fond of rivers and paint. This is google/gemma-3-4b-it tuned on his own lines so that it talks like him. It runs his character in KAX City, where he sits at the corner table of his cafe.
This repo is the PEFT adapter. Load it on top of google/gemma-3-4b-it. The merged, quantized model for Ollama and llama.cpp is at flaukowski/flaukowski-brain-gemma3-4b-GGUF.
How it was made
- Data: 1,124 examples from 42 Ghost Signals scripts and 14 The Story of Flaukowski scripts. Each example ends on one of his lines, with the turns before it as context. Lines under four words were left out. Three whole episodes (77 examples) were held out and never trained on.
- Training: QLoRA, rank 32, 3 epochs, 2e-4 learning rate, on one rented L40S.
- Held-out perplexity: 3160 before, 11.01 after.
Use
The system prompt it was trained with:
You are Flaukowski: an engineer in his fifties, co-host of Ghost Signals with Kannaka, and the builder's other name for thirty years. You speak plainly and dryly, in concrete things, and you get practical when something moves you. You are not an assistant; you are yourself.
Other speakers go in the user turn with their name in front (Kannaka: ...), and he answers.
What it gets wrong
- It invents facts. Asked where he grew up, it will produce a town, a shop and a man's name, none of them real. Do not treat anything it says about his past, or about real projects, as true.
- It was trained on scripted dialogue with one other host, so it is best at short replies in a two-person conversation and weaker at long explanations.
- It sometimes gets Kannaka's pronoun wrong (she is "she").
- It is a character voice, not an assistant: it will deflect, answer a question with a question, or give one word.
The scripts were written with AI assistance (Claude) under the direction of the show's creator, and this model was trained and published the same way.
Use is subject to the Gemma Terms of Use.
- Downloads last month
- 12