speleoalex/physisml-it-preview

PhysisML — Italian curriculum 0-12 (experimental)

A small language model trained from scratch on a developmental curriculum: it learns like a child, sounds first, then words, then sentences, guided by a tutor that adapts the material to what the model is currently failing at.

This is a research preview of an experiment, published so the results in the repository can be checked against real weights. It is not a general-purpose assistant and it will not behave like one.

• Code, curriculum and full documentation: https://github.com/speleoalex/PhysisML
• Architecture: decoder-only transformer, pre-LayerNorm, lm_head weight-tied to the token embedding
• 23.6M parameters — d_model=512, n_layers=6, n_heads=8, d_ff=2048
• Context window: 128 tokens
• Tokenizer: byte-level BPE, 9000 slots allocated, 2590 active (the vocabulary can grow during training; unused slots are masked to -inf at inference)
• float32, runs on a CPU

These weights are the end of the curriculum: the level-12 checkpoint of a completed 0-12 build — phonemes, syllables, words, grammar, dialogue, literature (Rodari, Manzoni, Dante), class membership, and asking about a name it has never met.

⚠️ Italian only — nothing else has been tested

Every result below was measured on Italian, and Italian is the only language this model has ever been trained or evaluated on.

The repository contains an English curriculum skeleton (training_files/en/), but no English run has been trained, no English checkpoint exists, and no English number has been measured. Prompting this model in any language other than Italian is untested territory — expect noise, not a translation.

The same caveat applies to the curriculum design itself: the level ladder (phonemes → syllables → words → grammar → dialogue → literature) was built around Italian phonotactics and Italian morphology. Whether the approach transfers to another language is an open question, not a claim of this release.

What it can do

It answers the kind of prompt its curriculum drilled — short Italian questions about a closed world of familiar nouns and verbs. Every line below is an actual greedy output of these weights, through the bundled generate.py:

di: il cane                            → il cane!
cos è un cane?                         → il cane è un animale.
dove dorme il cane?                    → il cane dorme in casa.
chi è più grande, il cane o il gatto?  → il cane è più grande del gatto.
di: cosa mangia il cane?               → il cane mangia il pane.
perché il cane mangia?                 → il cane mangia perché ha fame.
cosa ha mangiato il cane?              → il cane ha mangiato il pane.
cosa mangerà il cane domani?           → domani il cane mangerà il pane.
il pane è buono?                       → secondo me il pane è buono perché è caldo.
il cane è fedele?                      → secondo me il cane è fedele perché aspetta.
commenta il libro                      → il libro insegna, e questo è importante.
l albero è una pianta, questo è un tamburo  → cos è un tamburo?

The last line is level 12: shown a name it has never met, the model asks instead of inventing.

Results

Exact match against the curriculum's gold answers, greedy decoding, 849 graded prompts across the thirteen levels.

The published checkpoint against every level — one model, asked everything it was ever taught:

L0	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11	L12
100%	94%	80%	84%	88%	94%	92%	85%	87%	77%	95%	79%	100%

Mean 89.0%, self-repetition 1.4%.

For scale: in the project's previous build the final checkpoint scored 20% on the same question, and ten extra consolidation cycles brought it only to 48%. Here no level of the finished model sits below 77%.

The usual per-level table — each level scored on its own snapshot, which is a much easier question — averages 94.9% on this run:

L0	L1	L2	L3	L4	L5	L6	L7	L8	L9	L10	L11	L12
100%	89%	70%	89%	100%	99%	95%	100%	100%	100%	97%	94%	100%

Five things changed between the two builds — six dream cycles per level by default, target pools widened from 227 to 728, a vocabulary retrained without punctuation glued to words, one gold answer per prompt enforced across levels, and <|EOS|> registered again — so the gain cannot be attributed to any single one of them.

Where it fails

Level 11 teaches class membership, and the cos è X? form is asked about only 24 nouns. Eighteen more appear in the level's yes/no steps without ever being asked that way, and cane appears only as an answer (fai un esempio di animale). The consequences are visible:

cos è un cane?             → il cane è un animale.     ✓
fai un esempio di animale  → il cane è un animale.     ✓
il cane è un animale?      → sì, il cane è un animale. ✓
cos è il cane?             → il cane è una cosa.       ✗ (a superordinate, not the class)

The shape is right and the class is not: the model has the fact and no drilled path from that phrasing to it.

Level 12 teaches asking about a name it has never met, and it does that well in the sentence shape it was taught (l albero è una pianta, questo è un tamburo → cos è un tamburo?). A bare chi è zibaldone? was never taught, and there it confabulates — il fratello è una persona.

Beyond that:

• No world knowledge. It was trained on a synthetic teaching curriculum of a few megabytes, not on a web corpus. Anything factual it produces is invention.
• No instruction following beyond the prompt shapes in the curriculum.
• Closed vocabulary — 2590 active tokens. Out-of-curriculum words break it.
• 128-token context, so no documents, no long conversations.
• No alignment or safety tuning of any kind. There is no refusal behaviour, no filtering, no RLHF. It is a research artifact, not a product.
• Sampling costs a lot here. Every example on this page is greedy (--temperature 0). With sampling on, cos è un cane? answers animale, persona or luce depending on the draw: at this scale the model has the right answer at the top of the distribution, not alone in it.

Do not put this in front of users. Use it to study the training method.

How to run it

There is no AutoModel support: the architecture and the tokenizer are custom, so the model loads through the small package shipped in this repo.

pip install torch safetensors numpy
hf download speleoalex/physisml-it-preview --local-dir physisml-model
cd physisml-model

python3 generate.py "di: cosa mangia il cane?"   # one answer, greedy
python3 generate.py                              # interactive REPL
python3 generate.py --no-affect "di: il cane"    # plain transformer

Files:

File	What it is
model.safetensors	the weights, float32. lm_head.weight is absent on purpose — it is tied to tok_emb.weight and the loader re-ties it
config.json	architecture, active vocabulary size, context window
tokenizer.json	byte-level BPE vocabulary + merges
physisml/	inference code: model, tokenizer, sampling, affective modulation
generate.py	CLI: single prompt or REPL
MANIFEST.json	which checkpoint each artifact came from, with sha256

generate.py is a port of the repository's own generation path, so the affective system (confidence, pleasure, pain, fear shifting the logits at every step) is active by default — --no-affect turns it off if you want to see what the bare transformer does.

One detail it cannot reproduce: the published exact-match scores are measured with the punctuation stop threshold set from the length of the gold answer, which a free-form prompt does not know. Expect the model to run past its answer more often here than the tables suggest.

How it was trained

Two phases per level, repeated up the ladder:

1. Text phase — ordinary self-supervised training on the level's corpus.
2. Teaching phase — a tutor poses a prompt, grades the model's answer (+++ … -), and the feedback drives the update. The tutor picks the next prompt from what the model is currently getting wrong.

Each session ends in a dream: a consolidation pass that replays every level's question-answer corpus, with no new teaching. It is where cross-level retention comes from, and the effect is large enough to measure on a single cycle. Taken at level 6 of this same build, immediately before and after one dream, on all seven levels the model had seen:

	L0	L1	L2	L3	L4	L5	L6	mean
before	62%	0%	0%	0%	3%	1%	100%	23.7%
after	95%	91%	74%	74%	72%	83%	100%	84.3%

Teaching level 6 had erased levels 1 to 5 outright; one replay pass brought them back and cost the current level nothing. The build runs six such cycles per level.

The tutors are, in cost order: a rule-based deterministic teacher (offline), a hybrid teacher using a small local LLM for the grading (offline, via ollama), and optionally the Claude API. The whole Italian curriculum trains offline — every level ships its own local teacher configuration, so no API key is needed to reproduce this model from scratch (./build.sh 12, about two and a half days on a 16-core CPU).

Training data is the curriculum in training_files/it/: hand-built and script-expanded target pools, question-answer pairs, and level texts. Parts of the question-answer material were seeded by earlier teaching sessions that used the Claude API tutor.

License and attribution

MIT — Copyright (c) 2026 Alessandro Vernassa. See LICENSE.

@software{physisml,
  author = {Vernassa, Alessandro},
  title  = {PhysisML: a language model trained on a developmental curriculum},
  year   = {2026},
  url    = {https://github.com/speleoalex/PhysisML}
}

The name is φύσις (physis, nature, growth) + ML. The documentation exists in both English and Italian in the repository.