AI made by Sadik Nawaz, 11 years old, 6th class.
A 7.9M-parameter transformer trained from scratch on English, Hindi, Hinglish, code, maths and world knowledge, plus the complete local chat app it runs in. "Medium" is the project name: the weights are intentionally small so the whole thing runs offline on a home PC with no GPU.
A static demo runs this exact checkpoint locally with ONNX Runtime Web. No server, no API key, and the text you type stays in your browser:
https://huggingface.co/spaces/SadikNawazAI/sadik-ai-medium-multilingual
The demo is the raw base model continuing text. The chat app with lessons, memory and tools described below is separate application code, not part of these weights.
| field | value |
|---|---|
| creator | Sadik Nawaz (11, 6th class) |
| parameters | 7,866,048 (1,574,592 non-embedding) |
| training | 63.9 M tokens · step 4249 |
| validation loss / perplexity | 2.9338 / 18.8 |
model/ inference-only checkpoint + architecture JSON
tokenizer/ byte-level BPE tokenizer (32,768 vocab)
generate.py standalone text sampler (torch only)
join_weights.py rebuild weights.pt from split Hub parts (checks SHA-256)
app/ full chat app: memory, web, maths, code, media, lessons
MODEL_CARD.md what this is / what it can't do
DATA_SOURCES.md training sources + licenses
PUBLISH.md how to publish this folder (HF / GitHub)
SHA256SUMS checksum for every file in this release
requirements.txt
LICENSE.txt
pip install -r requirements.txt
python generate.py --say "hello" --max-new-tokens 80
The download zip already contains model/weights.pt, so nothing extra is
needed there. The Hugging Face repo stores the same file as eight small parts
(model/weights.part00.pt ... part07.pt) because long uploads fail on the
home connection used to publish it. generate.py and the chat app join the
parts automatically, or you can do it explicitly:
python join_weights.py # rebuilds model/weights.pt, verifies SHA-256
python join_weights.py --check # verify an existing file only
The part sizes and hashes live in model/weights.parts.json.
From the extracted folder root:
python app/chat_server.py --checkpoint model/weights.pt --tokenizer tokenizer/tokenizer.json --port 8000 --open
The app answers maths and common coding asks from its curated layer, keeps memory across sessions, looks facts up on the web, and can generate images, speech audio, and PDF/Word/Excel/CSV files. It also gives 113 offline lessons across English, Hindi, maths, science, history, geography, civics, computers, life skills, arts, sports and world knowledge, and writes short stories or poems from a curated library.
who made you? — the creator creditteach me about teamworktell me about world religionsexplain fractionswho was Ashoka?what is the Constitution of India?how does artificial intelligence learn?/learn Hinduismhow do I study better?what should I do for a small cut?write a story about a dogwrite a poem about the moon/topics — every offline lessonwhat is 12 x 13?code for fibonacci in pythondraw a red fox in the snowThe language model itself is a base next-token model: it continues text.
Stories and poems come from a small curated library, with the model as a
fallback. Facts, maths, coding answers and school lessons come from the app
layer, which is why the release is useful offline even though the network is
small. See MODEL_CARD.md for honest limits.
pip install huggingface_hub
hf auth login # older: huggingface-cli login
hf upload SadikNawazAI/sadik-ai-medium-multilingual . --repo-type model
# older: huggingface-cli upload SadikNawazAI/sadik-ai-medium-multilingual . --repo-type model
Everything here was produced by export_release.py; rebuild after retraining:
python export_release.py --checkpoint runs/medium/best.pt
model/weights.part00.pt 4 MBmodel/weights.part01.pt 4 MBmodel/weights.part02.pt 4 MBmodel/weights.part03.pt 4 MBmodel/weights.part04.pt 4 MBmodel/weights.part05.pt 4 MBmodel/weights.part06.pt 4 MBmodel/weights.part07.pt 4 MB