sudanzh

A small MiniGPT model trained with TAIT that answers questions about Sudan in Simplified Chinese.

Model details

  • Format: TAIT .npz (NumPy weights + BPE tokenizer, readable by the KAIT runner and by TAIT-TOY in the browser)
  • File: sudanzh-0.1.npz (11009118 bytes)
  • Vocab size: 1503
  • d_model: 256 | d_ff: 1024 | block_size: 57
  • TAIT version: 2.3.1
  • Training data: sudan-china-facts-zh (150 facts about Sudan, China and China–Sudan relations, 1363 examples)
  • Training examples: 1363
  • Epochs: 60
  • Final training loss: 0.9368
  • Validation loss: 2.9876

Honest evaluation

This is a small experimental model, and its limits are stated plainly. On 40 held-out questions in new phrasings (greedy decoding), it answered 5 correctly (12.5%). It memorises its exact training questions well but often garbles or gives wrong answers to new phrasings. It is a demonstration of training a Chinese Q&A model on weak hardware with TAIT, not a reliable source of facts. The TAIT-TOY web app shows its best verified answers as suggestion chips when the interface language is Simplified Chinese.

How to use it

pip install "tait[publish]"  # publish support; training itself needs only NumPy
# Download the model, then chat with it locally:
tait chat --model sudanzh-0.1.npz

You can also load the same .npz file in KAIT, the Android runner for TAIT models, or in TAIT-TOY in the browser, by pasting this repository link there.

Limitations

This is a tiny model trained from scratch on a small dataset. It may hallucinate, repeat itself, or answer off-topic. Treat its output as an experiment, not as a factual source.

Source

Trained with TAIT: https://github.com/muzammilsoft/tait Dataset: https://huggingface.co/datasets/muzammilsoft/sudan-china-facts-zh

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support