Why this matters
Large pre-trained models often refuse or redact outputs for safety reasons; researchers and creators sometimes need controlled, local access to models with those refusal paths altered for analysis, roleplay, or exploration. RVN demonstrates a reproducible, low-damage way to surgically remove refusal circuitry from a high-capacity Qwen3.8-27B model using repeated ARA (Arbitrary-Rank Ablation) passes, trading stricter guardrails for recoverable and measurable behavioral changes.
What Sets It Apart
- Double-refined ARA pipeline: RVN builds on trohrbaugh's heretic-ara and applies two additional full-weight ARA passes, pushing residual refusals lower while keeping behavioral damage minimal. Measured refusal rates: 0–1/100; KL vs base ≈ 0.0085.
- Metrics-first preservation: ARA optimizes weight matrices to Preserve (low KL), Steer (pull harmful activations toward good-prompt manifolds), and Overcorrect (push away from original harmful outputs), giving finer control than single-direction subtraction.
- Deployment-ready quant spectrum and formats: published as GGUF files across many quants (Q8_0 → IQ1), with multilingual, MTP (speculative-decoding head) and vision-compatible variants, sized for a range of GPUs and llama.cpp workflows.
- Clear licensing and provenance: base model and abliteration sources retain Apache-2.0 licensing; authorship and calibration receipts are documented in the model card.
Who it's for and tradeoffs
Great fit if you need reproducible, local experiments that require fewer automated refusals (researchers analyzing model safety/pathology, creative writers or roleplayers testing uncensored outputs, or developers prototyping offline LLM-powered features). RVN makes it practical to run Qwen3.8-27B-class models on consumer GPUs via multiple quant options and embedded MTP for speed.
Look elsewhere if you need a production-grade, safety-hardened model for end users, minors, or regulated environments: RVN intentionally reduces some safety guardrails and retains only a subset of the strongest refusals. Behavior can vary across domains and languages; treat outputs as unmoderated and use responsible policies and post-processing when deploying.
Where it fits
RVN is positioned between the original Qwen3.8-27B (base) and trohrbaugh's heretic-ara variant: it prioritizes aggressive but measured removal of refusal behavior while minimizing KL damage. Use RVN for local, offline research and experimentation; prefer upstream, safety-preserving variants for public-facing services.
