Reload what you saved.
The saved reading profiles reproduced 512 heldout answers after independent cold reload. The starter carries their inference programs and tokenizers; the base remains separate.
What the asset keeps ↗ARTI A place for your computation
Your ideas.In the model.
Give a model a computation of your own. Compose its operations, choose what learns, and work with its state in a program you can read and run.
Start on CPU. Add one operation. Keep its declaration.
A neural runtime for PythonPyTorch / JAX3.3.0
A computation, in inkARTI / FormulaScan
xₜsequence input
hₜ₋₁incoming carry
Learnable valuesgain · retention
σgate
tanhcandidate value
hₜnext carry
Σ haccumulated carry
Shared declaration
15 typed instructions. One bounded recurrence, shared by PyTorch and native JAX.
Make it your own
One intention. A different response.
Suppose you want projected negative features to become gentler before the next projection. Keep the inputs and values, add SiLU to the Formula, and make that behavior part of the declaration.
LoRA is one Formula program. The scalar illustration explains the added operation; whether it helps a task requires a separate experiment.
This time, let the projected features pass through SiLU before the next projection.
Your extra operation: SiLU
base + B(SiLU(Ax))
5 instructions65,536 unchanged valuesA scalar view of this operation: −2 becomes about −0.238. The response changes before B.
Try the LoRA → SiLU exerciseGrouped declarations on fixed neural-layer inputs, gain = 1, with no training updates. This is a computation edit, not improved reading quality.
Your first piece of work
Keep the change. Read it again.
The first exercise starts with LoRA, adds SiLU, then saves and reloads the changed declaration. You leave with a readable program and a record of what changed.
- 01Take one small script
- 02Run it on your computer
- 03Keep the declaration and record
Python · CPU dependencies. No base model or training. The exercise reloads with the same caller-supplied bindings.
The operation you added
SiLU{"mode":"silu"}The same output position ·
y[0,0,0]0.2635770.212425
Fixed inputs · before / after adding SiLU · recorded values shown to six decimals
- Computation
- 4 → 5instructions
- Bound values
- 65,536unchanged
The state you pass inh₀
5 recurrent steps [0, 0, 0]The state returned to youh₅
[0.170, 0.301, -0.119]0102030405
h[0]h[1]h[2]
A state that comes back to you
Your start. Your next state.
In the public recurrence, you pass in an initial state. Five steps later, the program returns a new one while leaving your original unchanged. You decide how the next computation continues.
Follow the stateThese are five recorded state transitions, not training updates or a reading-quality result. PyTorch and bounded native JAX execute the shared declaration.
A saved result you can take with you
Now bring your own material.
Take the saved English or Chinese reading program. Begin with its small example, then try a short policy, note or passage of your own on your computer. This is an entry point, not a claim of model superiority.
No training or GPU. First base-model download ≈ 1.2 GB. This small model can misread a passage; check its answer against the source.
Your passage
Orders placed before 2 pm ship the same day. Standard shipping takes 3 business days. Returns are accepted within 30 days of delivery.Your question
How many days after delivery can an item be returned?
Generated answer30
Actual output from the program included in the starter. Check answers to your own material against its source.
04 Keep the work
An experiment ends. The work can continue.
A declaration you can inspect. An asset you can reload. A supported training state you can return to.
Continue a supported run.
In each language, the three-update save/reload continuation matched parameters and data cursor. Inference assets and continuation checkpoints serve separate purposes.
Understand recovery ↗Compare the same computation.
Standard LoRA can be a Formula. The paired check keeps the formula, trained weights, mounting and precision aligned; broader program choices need their own experiments.
Inspect the paired check ↗Where would you like to begin?
A small first step. Room for your next idea.
Open implementation. Your next step.
python -m pip install "arti-fit==3.3.0"