ARTI guides
Reading QA on your CPU
Load a saved English or Chinese program, answer a short passage, then paste your own material. No training or GPU is required. The first model download is about 1.2 GB.
- 01Get the package5.2 MB
- 02Install oncePython · CPU
- 03Ask a questionA recorded example
- 04Use your materialPaste and ask again
1. Get the ready-to-run package.
A saved English program, a saved Chinese program, and one script. Extract the ZIP, then open a terminal in the arti-reading-starter folder.
Download starter · 5.2 MB ↓Before you download
Use Python 3.10+ and CPU; no training or GPU is required. Allow about 1.2 GB for the first base-model download, plus Python dependencies. We suggest a computer with at least 8 GB RAM.
Plan the first download and wait
For the 1.2 GB base alone, a sustained 20 Mbps transfer takes about 8 minutes; 100 Mbps takes about 2 minutes. These are file-size calculations, excluding dependencies, server limits, installation, model loading and answers.
We do not yet have a public loading or answer-time range across devices. The terminal identifies downloading, loading and generating; interactive questions reuse the loaded model.
A first run for model developers
Run a saved computation before building your own. Qwen reads the text; ARTI reconstructs the attached program from its declaration and learned values. Continue to model integration when you want to define a different component.
See how a component attaches ↗2. Install the CPU environment.
Run these commands inside the extracted folder. They create an isolated environment and install CPU Torch plus the pinned runtime dependencies.
python -m venv .venv
.venv\Scripts\python -m pip install "torch==2.11.0" --index-url https://download.pytorch.org/whl/cpu
.venv\Scripts\python -m pip install -r requirements.txt3. Ask the first question.
.venv\Scripts\python ask.py --language enThe first run downloads the public base from Hugging Face. The script then loads the saved ARTI program on CPU and answers the included question. It prints the answer and writes answers.jsonl in the starter folder.
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.
Success means the terminal shows a generated answer and an answers.jsonl file exists. Later runs reuse model-cache/.
4. Paste your material and ask again.
.venv\Scripts\python ask.py --language en --interactive- Paste a short passage, then enter an empty line.
- Type a question about that passage. You can ask more questions without reloading the model.
- Enter /quit to exit. Restart the command to use a different passage.
Start with roughly a hundred words. The script checks the complete input and answer budget and asks you to shorten overlong material. Keep the passage for checking the answer.
Where ARTI enters this result.
Qwen reads the text. The saved ARTI nonlinear program is attached after its last two MLPs. The package retains that program configuration, 131,072 learned values and tokenizer identity. The public predictor reconstructs this supported profile from its manifest.
This starter selects one nonlinear residual program. Standard LoRA can also be expressed as an ARTI Formula. ARTI provides the program vocabulary for composing these updates, routing and explicit state; the result depends on the computation and training you choose.
For general reading QA, you can use an existing model interface directly. Choose ARTI when your work involves defining and retaining learned computation inside the model. This example lets you inspect that program and its asset.
Try one change: add SiLU to LoRA.
Add one nonlinear node between the low-rank projections. Keep the input interface, the A/B values and the executor. Save the new declaration as JSON and load it again. This exercise needs the CPU environment above; it downloads no base model.
Before
base + B(Ax)After · gain = 1
base + B(SiLU(Ax))hidden = api.scalar_map(hidden, mode="silu")Download the complete exercise ↓ Save it in the starter folder, then run:
.venv\Scripts\python change_lora_program.py- First run the script as downloaded. It saves program-change/program.json and result.json.
- Open the Python file. Inside with_nonlinearity(), change mode='silu' to mode='tanh', then run the same command again.
- Look for same_bindings: true, five instructions, a positive output change and reload_max_abs_difference: 0. The saved declaration now contains tanh.
The recorded run uses four fixed 1024-dimensional neural-layer inputs and 65,536 unchanged Bank values. The program grows from 4 to 5 instructions; its output changes. Reloading the saved declaration reproduces the changed output exactly.
This is a neural-component editing exercise with zero training updates. Its initialized values and output change are not reading-quality results. A new task program needs its own training and evaluation.
Inspect the recorded SiLU result on this page
{
"scope": "Fixed neural-layer computation edit; zero training updates",
"input_shape": [
1,
4,
1024
],
"rank": 32,
"bank_values": 65536,
"before_instructions": 4,
"after_instructions": 5,
"same_bindings": true,
"output_max_abs_change": 0.1494123935699463,
"reload_max_abs_difference": 0,
"script_sha256": "b8ce7c4788fae07480c0187326e68594f7f41b3cfe9b944df9231995f16b9628",
"package": "arti-fit==3.3.0"
}Choose what to do next.
- For your own passages: keep using the local script.
- To change the computation: follow model integration and the runtime guide.
- To train a new reading profile: use the full English/Chinese training case.
This guide targets arti-fit 3.3.0 · Public source dcb321d6