Skip to content

coba:400m-fp16

16 TagsUpdated 421M params512 contextEnglishApache-2.0by Ollaya

Ettin 400M: 0.591 on typed decisions, zero-shot; 24 ms for five questions on an RTX 5090.

421m
ollaya run coba:400m-fp16 --preset triage "I was charged twice for my subscription this month and want a refund."

Details

  • graph49c15752c4a3 · 3 MBonnx · encoder · 421M · fp32
  • graph27779f73da3c · 4 MBonnx · encoder · 421M · fp16
  • weights90a4b8c3f1c3 · 1.7 GBhuggingface.co/ollaya-dev/coba-400m/resolve/b9db262…/model.safetensors
  • tokenizer6c8aaa9a5420 · 4 MBhuggingface.co/ollaya-dev/coba-400m/resolve/b9db262…/tokenizer/tokenizer.json
  • decisionb01dafbc1f88 · 457 B{"engine": "onnx", "family": "laya", "encoder": "encoder", "layout": "laya-markers-v1", …}
  • calibrationa234821048b8 · 87 B{"temperature": [0.91, 0.91, 0.55]}
  • licenseef14791048a7 · 10 KBCoba by Ollaya (https://huggingface.co/ollaya-dev)
  • arch8be02e5f6a3b · 1 KBapplication/vnd.ollaya.arch
  • paramsa1d447ba672f · 21 Bprecision fp16

Every layer is checked against its sha256 when it is pulled. Weights and tokenizers download from the model author's Hugging Face repository at a pinned commit; Ollaya never re-hosts them.

Variant of

coba readme and all models