nicu-5m

Nicu 5M — a Sicilian character AI that runs in your browser

Nicu is a 5M-parameter language model trained from scratch (no pretrained base) that plays a single character: a friendly guy from Catania, Sicily. He knows his small world — the sea, the grill, his friends, Sant'Agata — and deflects instead of hallucinating on everything else. That's the design: a model this small can't hold facts about the world, so it stays in character and dodges, usually with a joke.

Try it live: nicu.chat — the model runs entirely on your device (ONNX Runtime Web, WASM). No inference server.

This is the v1.0 release, part of a three-size family trained on the same final corpus for size-scaling comparison: nicu-5m · nicu-9m · nicu-20m

How it was made

A large teacher LLM generated ~565k synthetic in-character dialogues from a "character sheet" (identity, voice, boundaries). After quality filtering, a ByteLevel-BPE tokenizer (vocab 6000) and this nanoGPT-style decoder were trained from random initialization on the resulting 38M tokens.

Architecture

Evaluation

Blind LLM-judge over a 204-case character-fidelity suite (identity, substance, multi-turn coherence, boundary handling): 59.3% case pass rate. The int8 ONNX was judged blind against fp32 on the full suite: no measurable quality loss (the live site serves int8).

Files

• model.onnx — fp32 ONNX export (opset 17), for onnxruntime-web/onnxruntime
• model.int8.onnx — dynamic-int8 quantized ONNX (what nicu.chat serves)
• tokenizer.json — ByteLevel BPE tokenizer (HF tokenizers format)
• checkpoint.pt — PyTorch weights + config ({'model_state', 'cfg', ...})

Prompt format

Plain dialogue text, no special tokens beyond <|endoftext|> (id 0):

Utente: Ciao Nicu! Chi sei?
Nicu:

Sample with repetition penalty ~1.15 (current turn only), temperature 1.0, top-p 0.92; stop on <|endoftext|> or on a generated \nUtente: / \nNicu: line.

Versions

• v1.0 (current) — retrained from scratch on the final corpus; family of three sizes (5M / 9M / 20M); int8 ONNX included.
• v0.9 — first public release (20M only).

Limitations

• Italian only, heavily flavored with Catania dialect and running gags.
• Not factual, not a task assistant — an entertainment character.
• Out-of-domain questions get in-character deflections by design.
• Short conversational memory (context 512); factual recall of user details across turns is weak in this generation.

Links

• Live demo: https://nicu.chat
• Code (browser runtime + export): https://github.com/Federico-Anastasi/nicu-chat
• Author: @FedeAnastasi

License

Weights: CC-BY-NC-4.0 (non-commercial). Contact the author for other uses.