The full timeline

Essay arguing users should choose their own model alignment and documenting fine-tuning-based uncensored WizardLM models.

Report undoing Llama 2-Chat 70B safety training with LoRA for under $200 on one GPU.

Precursor paper establishing that safety fine-tuning is not robust when weights are public; later at ICLR 2024 workshop.

Introduces Contrastive Activation Addition, computing steering vectors from residual-stream activation differences; direct precursor to refusal-direction work.

Informal write-up previewing the paper; states refusal is mediated by a single residual-stream direction.

Among the earliest dated FailSpy Llama-3 abliterations, per HF upload commits.

Reddit post explaining 'ablated + obliterated = abliterated'; first documented public use as a method name; also documents Phi-3 abliterations.

Abliterated Daredevil-8B healed with DPO to recover benchmark performance.

Pure HF Transformers proof-of-concept for refusal removal; base for deccp and huihui-ai.

Qwen2-7B-Instruct-deccp plus deccp dataset, built on Sumandora's code; reduces refusals from ~100% to ~20%.

Hugging Face blog tutorial (proofread by FailSpy) that popularizes abliteration with a Colab notebook.

Demonstrates the finding across 13 open chat models up to 72B and proposes a rank-one weight edit disabling refusal.

CAA appears in ACL 2024 proceedings, pages 15504-15522.

Extends single-direction account; argues refusal is an affine function of activations.

Refusal-direction paper appears in NeurIPS 2024 proceedings.

Early huihui-ai abliteration built on Sumandora's code; start of high-volume publishing program.

Python library for layer ablation/addition, built on Sumandora's repo; Show HN January 2025.

Abliterated DeepSeek-R1 distilled models after R1's 20 Jan 2025 release.

Gradient-based approach revealing multiple independent refusal directions and concept cones.

Extended-refusal fine-tuning keeps refusal-rate drop under abliteration to at most 10% vs 70-80% baseline.

Wollschläger et al. paper presented at ICML 2025.

HF discussion: 'we are already trying, it's a bit challenging, and the project is still ongoing.'

Evaluates which data-centric safety components survive abliteration across SmolLM2-1.7B checkpoints; NeurIPS 2025 Lock-LLM workshop.

Publishes Huihui-Kimi-K2-Instruct-0905-BF16-abliterated-GGUF (1T total / 32B activated params).

Automatic abliteration tool combining directional ablation with TPE/Optuna parameter optimization.

Evaluates Heretic, DECCP, ErisForge, FailSpy across 16 models; tool compatibility 16/11/9/5.

Revised version submitted to arXiv.

Adds LoRA engine with 4-bit quantization support.

Dated blog coverage reports Heretic as #1 trending repository.

SVD-based abliteration toolkit with Gradio UI, CLI, Python API and telemetry; 1,000 stars in 24 hours.

FT/Alice investigation: Llama 3.3 stripped in under 10 minutes; Gemma 4 within 90 minutes of release.

Reports Heretic produced 3,500+ decensored models downloaded 13 million times (per Weidmann to FT).

CTC Sentinel Vol.19 Issue 7 commentary by Adam Hadley on the open-weight abliteration problem.

Kimi K3 served from 16 July 2026 per West Point CTC commentary.

6,644 distinct guardrail-free HF models downloaded 22M+ times in 30 days; via Business Wire, Forbes.

Open weights published 27 July 2026 following UK AISI / US CAISI evaluation.