Sleeping
Agents
Apoha Bge Small En V1.5
📈
Classify text into banking or medical concepts with OOD safety
EVAM Research Lab
non-profit
AI & ML interests
None defined yet.
Recent Activity
Organization Card
EVAM Lab 🪷
Ancient logic. Modern AI. Built to know what it doesn't know.
We implement Buddhist formal logic as ML infrastructure — rigorous epistemological frameworks developed over 1,500 years, now running on GPUs.
Phase A: Apoha — Available Now
"A concept is not defined by what it includes, but by what it excludes." — Dignāga, c. 480–540 CE
Apoha is a few-shot classifier that defines concepts by their exclusion sets — examples of what they are NOT. Inputs that don't clearly belong anywhere receive UNCERTAIN instead of a forced label. OOD rejection is built into the architecture, not bolted on after.
| Benchmark | Apoha OOD Rejection | Best Baseline | Improvement |
|---|---|---|---|
| CLINC150 (30 intents) | 84.6% | 15.1% (InfoNCE) | 5.6× |
| Banking77 (50 intents) | 29.1% | 0.0% (InfoNCE) | ∞ |
| HWU64 (40 intents) | 24.9% | 0.2% (InfoNCE) | 166× |
| Cybersecurity (15 TTPs) | 45.2% | 1.2% (InfoNCE) | 38× |
👉**apoha-bge-small-en-v1.5** — model weights + inference code
👉 apoha-demo — live interactive demo
📄 Preprint: ArXiv (coming soon)
Roadmap
| Phase | Module | What it does |
|---|---|---|
| A ✅ | Apoha | Rejects OOD inputs — classifier knows what it doesn't know |
| B 🔨 | Hetuchakra | Verifies LLM outputs — formal argument gate |
datasets 0
None public yet