decima
7 TagsUpdated Apache-2.0by A. M. Madani
A. M. Madani's Decima: multilingual encoders (mmBERT-base, multilingual-e5-small) with a late-interaction scorer that reads every option against the state, so option order never changes the answer, and an ordinal head for scores. A general model, one for coding-agent decisions, two guards for LLM apps, and a small one that is the fastest on a CPU.
decisionmultilingualfine-tuned122m140m321m
Tags
| Name |
|---|
| decimalatestSame as decima:base.1.3 GB · 512 ctx · 100+ languages |
| decima:agentDecima-agent 2.3: decima-base fine-tuned for coding-agent decisions (secret and command gates, tool and model choice) and knowledge-graph judgments; states up to 8,192 tokens, options kept whole beside long questions.1.3 GB · 8192 ctx · 100+ languages |
| decima:baseDecima-base 2.0 (mmBERT-base, 321M), fp32: 0.483 on typed decisions; 27 ms for five questions on an RTX 5090.1.3 GB · 512 ctx · 100+ languages |
| decima:base-cleanDecima-base-clean 1.0: decima-base retrained without non-commercial data, for commercial use; 0.474 on typed decisions.1.3 GB · 512 ctx · 100+ languages |
| decima:guardDecima-guard 1.0 preview (mmBERT-base, 321M): prompt injection, jailbreaks, leaks and harmful content in user messages, tool results and model replies (OWASP Top 10 for LLM applications). English, Arabic and Persian.1.3 GB · 2048 ctx · 100+ languages |
| decima:guard-smallDecima-guard-small 1.0 preview (mmBERT-small, 140M): the same checks, faster, a few points lower.620 MB · 2048 ctx · 100+ languages |
| decima:smallDecima-small 1.1 (multilingual-e5-small, 122M), fp32: 0.414 on typed decisions; 591 ms for five questions on a CPU.510 MB · 512 ctx · 100+ languages |
Tags without a suffix load fp16 on a CUDA GPU and fp32 on CPU; add -fp16 or -fp32 to pin one precision.