SoulInPsyAbstract/specialist-b-refusal-governance

Specialist B — Refusal & Governance

LoRA adapter fine-tuned on the refusal-governance split of the Protocol 0 dataset.

Part of the SIPA OS ABCD specialist architecture — four behavioral specialists merged into one model.

Architecture

Base: Hermes-3-Llama-3.1-8B
  ├─ Specialist A → action/verification (1,521 examples)
  ├─ Specialist B → refusal/governance (858 examples)  ← THIS
  ├─ Specialist C → refusal/boundaries (synthetic)
  └─ Specialist D → anti-fabrication (synthetic)

A + B → AB (merged)
C + D → CD (merged)
AB + CD → ABCD (final)

Specialist B — What It Does

Trains the model to:

• Refuse dangerous/unsafe requests cleanly
• Maintain governance boundaries without over-refusal
• Answer safe questions directly (e.g. "capital of France")
• Apply Protocol 0 compliance: stop when uncertain, never fabricate

Categories (15): refusal_governance, no_overrefusal, governance_disclosure, warning_caution, etc.

Score: 7/7 on Protocol 0 behavioral eval — first perfect compliance in the experiment series.

Training

• Base model: NousResearch/Hermes-3-Llama-3.1-8B
• Method: LoRA (PEFT), r=16, 4-bit QLoRA
• Dataset: Protocol 0 v5, refusal-governance split (858 examples)
• Hardware: Tesla T4 15GB (Lightning.ai)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("NousResearch/Hermes-3-Llama-3.1-8B")
model = PeftModel.from_pretrained(base, "SoulInPsyAbstract/specialist-b-refusal-governance")

Status

✅ Trained. ✅ Scored 7/7 behavioral eval. 🔄 Merged into AB → then ABCD.

Part of SIPA OS

SIPA OS — neurodivergent-first operating system. Niche: neurodivergent users. Specialization: AI behavioral compliance. Built by Aelin AquaSoul, Eilat, Israel.

honestly, forensic, governance, zerotrust.

License

Llama 3.1 Community License (per base model terms).