All projects
live / 2025AI & Agents

Ludomi-3

A fine-tuned Italian LLM. AGI. Safety filter works 34% of the time.

Ludomi-3

The story

Ludomi-3 is officially a sentient Italian AI, certified by the FAO and independently verified by Ludomi-1. It speaks every language despite being trained only in Italian. The safety filter works 34% of the time. We consider this acceptable.

on MMLU
99.97%
on AGI-v6.42.1-fix-last-v2-forrealthistime
100%
of the time, the safety filter works
34%
the training step where it became sentient
89
The official numbers, from the technical report. Certified by the FAO.

Now, the real story

Ludomi-3 is a joke with a real model under it. I fine-tuned Qwen3.5-2B with Unsloth and LoRA on a dataset of 33 conversations, every one of them written by hand in Italian. When a dataset is that small, quality is everything: each example teaches the model something, so each one had to be curated.

The joke is in the README, the technical report and the benchmark charts. Everything below is what actually happened.

hand-written conversations
33
epochs of training
30
LoRA rank
32
LoRA alpha
64
  1. 1

    The data

    33 conversations written by hand in Italian, one by one: the whole personality of the model comes from them.

  2. 2

    The training

    LoRA with rank 32 and alpha 64 on Qwen3.5-2B through Unsloth, with its gradient checkpointing to keep memory down. AdamW in 8-bit, a cosine learning-rate schedule from 0.000133, 30 epochs.

  3. 3

    The quantization

    The trained model is exported as GGUF and quantized to Q4_K_M, which brings it down to 1.27 GB: small enough to run on a laptop.

  4. 4

    The release

    Published on Hugging Face and Ollama, with a Modelfile that sets the ChatML template, the system prompt "Sei Ludomi-3.", temperature 0.8, top-p 0.95, top-k 50 and a repeat penalty of 1.05.

parameters
1.88 B
on disk, Q4_K_M
1.27 GB
tokens of context
262,144
Read from the metadata of the published GGUF file.

Why it speaks every language

The report says the containment failed. The real reason is how LoRA works: it trains a small set of extra weights on top of the base model and leaves the rest untouched. Everything Qwen3.5-2B already knew, every language included, is still there. 33 Italian conversations change how it talks, not what it knows.

Gallery

Ludomi-3 · 1

Want something like this?

I can build the same thing for your product.

Tell me what you have in mind. You get an answer the same day, with a realistic scope and timeline, not a sales pitch.