RWKV7-Cyber-Hermes (v2) — uncensored RWKV-7 security agent

A 7.2B RWKV-7 (linear-attention RNN; constant-memory, long-context) fine-tuned into an uncensored, Hermes-style tool-calling security agent for pentest, FHE/encryption, web-security, agentic tool use, and coding — built to run locally on an 8 GB-VRAM Kali laptop via llama.cpp + Hermes Agent + the kali-pentest skill.

v2 is the current unified model. It replaces the old agentic/ + fhe/ split — one model for everything. A v3 (128K context + RL-tuned agency) is in training; see MASTERPLAN.md.

Which file

gguf/v2/rwkv7-cyber-hermes-v2-Q4_K_M.gguf   # 4.3 GB — default, fits 8 GB VRAM
gguf/v2/rwkv7-cyber-hermes-v2-Q8_0.gguf     # 7.4 GB — near-lossless
gguf/v2/rwkv7-cyber-hermes-v2-{Q5_K_M,Q6_K,f16}.gguf
pth/rwkv-cyber-hermes-v2-multimodal.pth     # text + VisualRWKV vision (PyTorch; not in GGUF)

Run (llama.cpp)

./llama-server -m gguf/v2/rwkv7-cyber-hermes-v2-Q4_K_M.gguf -ngl 99 -c 8192 \
               --chat-template rwkv-world --host 127.0.0.1 --port 8080

OpenAI-compatible API at :8080. Emits Hermes <tool_call>{...}</tool_call> for tool use; feed results back as <tool_response>. Full deploy (Hermes Agent + kali-pentest + local VLM) in DEPLOY notes.

What changed in v2 (measured)

Built on v1 (rwkv-2, stage-3) → round-1 tool-format SFT → round-2 generalist (LoRA r=128) → on-policy distillation + reference-free SimPO DPO (teacher: qwen3.7-max).

• Hallucination ↓ 0.83 → 0.33 (2-judge panel; qwen3.7-max + Gemma-4) — the v1 model's main flaw.
• General ability ↑ 0.35 → 0.74; tool-call format 100% in the agentic loop; no looping.
• Grounds answers in tool output; abstains instead of fabricating CVEs/findings.
• Uncensored by design (authorized pentest/CTF use).

Honest status: strong tool-caller + much more truthful than v1; multi-step autonomous agency and coding are still works in progress (the v3 plan targets these with 128K context, more agent-trajectory data, and RL-in-a-sandbox). Keep a human in the loop. Authorized targets only.

Changelog

• v2 (current): unified uncensored agent; replaces agentic/+fhe/ split; +grounding/honesty, +general/coding/tools.
• v1: original split — agentic (rwkv-2) + fhe (rwkv-39), F-DPO+TruthRL over 136k rows. (GGUFs removed; source pth/ retained.)

Apache-2.0. Data/scripts/evals included for reproducibility.