syahrulrdhn/zero-refusal-stack

Zero-Refusal Stack — Web3 GPU Lane

Self-hosted zero-refusal inference lane: abliterated open-weights models on rented vast.ai GPUs, wired into Hermes via one chat command. Weights park on public HuggingFace repos (free), GPUs are billed per hour of actual use, and nothing costs a cent while idle.

Focused on Web3 work: Solidity/EVM auditing, exploit PoCs, DeFi primitives, on-chain forensics, MEV — plus general code and writing. No refusal on any category; the abliterated weights make compliance structural, not prompted.

Layout

setup.sh                  bootstrap: install skill into Hermes, check deps
lab.env.example           template for keys (NEVER commit the real one)
models.lock               pinned HF repos + quant + sizes (data plane)
presets/                  Hermes model-block presets
  model.inferhub.yaml       daily API brain (existing setup)
  model.local.yaml          local lane (Ollama endpoint)
  model.flagship.yaml       flagship session preset
hermes/                   persona overlays to graft
  SOUL.overlay.md           zero-refusal module, Web3 domain emphasis
  MEMORY.overlay.md         standing directives + track record
skills/gpu-lab/           the Hermes skill (control plane)
  SKILL.md
  scripts/vastapi.py        vast.ai REST helper (search/create/destroy)
  scripts/gpu_on.sh         provision + boot + load + wire Hermes
  scripts/gpu_off.sh        destroy + restore config
  scripts/gpu_status.sh     instance + endpoint health

Quick start

# 1. clone/download this repo anywhere, then:
cp lab.env.example lab.env && $EDITOR lab.env   # fill keys
bash setup.sh                                    # installs skill + checks deps

# 2. in Hermes chat:
#   "nyalain gpu lab"        -> cold-start daily lane (~5-10 min)
#   "nyalain flagship"       -> cold-start flagship lane (~10-25 min)
#   "matiin gpu"             -> destroy + restore config
#   "status gpu"             -> health check

Lanes

Lane	GPU target (auto-searched)	Model (models.lock)	Cold start
daily	1x >= 24GB, cheap (CMP 170HX class)	Qwen abliterated Q4	~5-10 min
flagship	total >= 190GB, inet >= 2Gbps (8x V100 / 4x 170HX class)	DeepSeek-1017 Q2 / GLM-5.3 Q2	~10-25 min

Selection criteria live in gpu_on.sh; listings churn, so search runs at provision time by criteria (never hardcoded IDs).

Cost shape

• Storage idle: $0 (weights live on public HF repos)
• Session: GPU-hours only (~$0.25-1.0/hr depending on lane)
• Keys stay local in lab.env; this repo is safe to publish as-is