ARTI reading starter
====================

Run a previously trained public ARTI reading program on CPU. Training and
a GPU are not required. The first run downloads the external Qwen3-0.6B
base (about 1.2 GB); later runs reuse model-cache/.

Python 3.10+. Create an environment and install CPU PyTorch 2.11.0, then
the pins in requirements.txt. See https://arti.fit/docs/getting-started/.

python ask.py --language en
python ask.py --language zh
python ask.py --language en --interactive
python ask.py --language zh --interactive

In interactive mode, paste a short passage and finish it with an empty
line. Ask questions about that passage. Enter /quit to exit. To use a
different passage, restart the interactive command. Answers are also
written to answers.jsonl. The total prompt and answer budget is 384 tokens.

For a UTF-8 text file:
python ask.py --language en --material my-passage.txt --question "Your question"

The shipped examples are execution observations on owned example text,
not a new accuracy benchmark. This small model can return an incorrect
or incomplete answer; compare its output with the passage. Shipping vs
delivery wording produced a miss during qualification. The English/Chinese
benchmarks, sample eligibility and all comparison results remain published
separately at https://arti.fit/benchmarks/.

Programs: public SQuAD 1.1 / CMRC2018 profiles, seed 17, final step 200,
131,072 trainable values per saved nonlinear Formula adapter. See
program-sources.json and each artifact's manifest. No frozen host weights,
training checkpoints, optimizer state or original corpus are included.

Sources and licences
--------------------
ARTI runtime and public Formula implementation: MIT,
https://github.com/ragnarok-io/ARTI/tree/dcb321d6e108d0167c5fbca0a77501f34b154af2
Qwen3-0.6B and its tokenizer: Apache-2.0; QWEN-LICENSE.txt is included.
https://huggingface.co/Qwen/Qwen3-0.6B/tree/c1899de289a04d12100db370d81485cdf75e47ca
English dataset: https://rajpurkar.github.io/SQuAD-explorer/
Chinese dataset: https://github.com/ymcui/cmrc2018
Reproduction runner and protocol: https://arti.fit/benchmarks/
