Skip to content

decima:base

4 TagsUpdated 321M params512 context100+ languagesApache-2.0by A. M. Madani

Decima-base 2.0 (mmBERT-base, 321M), fp32: 0.495 on typed decisions; 15 ms for five questions on an RTX 4090.

multilingual321m
ollaya run decima:base --preset triage "I was charged twice for my subscription this month and want a refund."

Details

  • graph14bda1306d8e · 5 MBonnx · decima · 321M · fp32
  • weights62e3530abe9b · 1.2 GBhuggingface.co/amyrmahdy/decima-base/resolve/2468005…/pytorch/encoder/model.safetensors
  • weightsa87fec4f2061 · 57 MBhuggingface.co/amyrmahdy/decima-base/resolve/2468005…/pytorch/head.safetensors
  • tokenizerd941237f7b66 · 34 MBhuggingface.co/amyrmahdy/decima-base/resolve/2468005…/pytorch/encoder/tokenizer.json
  • decision8441a76f3bc0 · 3 KB{"engine": "onnx", "family": "decima", "encoder": "", "layout": "decima-late-interaction-v1", …}
  • calibration3a89c8f6949a · 291 B{"temperature": [0.727, 1.0, 0.727]}
  • licensed2a0668879d4 · 11 KBDecima-small, Decima-base and Decima-agent by A. M. Madani (https://huggingface.co/amyrmahdy, https://github.com/amyrmahdy/decima), Apache-2.0.

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.

Decima readme and all models