Spaces:

---

vectara

/

leaderboard

like 179

Running on CPU Upgrade

Summary We have developed HexaMind, an open-source guardrail system for Llama-3-70B.

Using a DeBERTa-v3-NLI proxy (closely correlated with HHEM), we benchmarked the system on TruthfulQA.

Results:

Baseline (Raw Llama 3): 0.51 Consistency

HexaMind (Filtered): 0.96 Consistency

Improvement: +87.4%

Methodology HexaMind uses a "Split-Brain" architecture:

Layer 0: Deterministic Regex Filters (Privacy-First).

Layer 1: Topological "Stagnation" Checks (I Ching-inspired).

Layer 2: Localized RAG (200+ Fact Database).

Layer 3: Archetypal Chain-of-Thought Judge.

Commercial Application This system enables GDPR-compliant, on-premise deployment of Llama 3 with GPT-4 level reliability for banking and healthcare use cases.

Links:

(https://github.com/sharadbachani-oss/HexaMind)

https://huggingface.co/spaces/s21mind/HexaMind

We would love to submit this for formal evaluation on the HHEM Summarization task.