What problem it solves

Refusal mechanisms in transformer models can block legitimate research, creative writing, and red-teaming alongside genuinely harmful content. Researchers need transparent, reproducible ways to locate refusal representations, evaluate their relationship to other capabilities, and apply controlled interventions without retraining or fine-tuning.

Key features

• Refusal direction extraction — Locate refusal-related representations through multiple extraction strategies, including PCA, mean-difference, sparse autoencoder decomposition, and whitened SVD.

  Source: README

• Analysis-informed interventions — Use analysis of refusal geometry and capability entanglement to configure the obliteration strategy, including target layers, direction count, and refinement passes.

  Source: README

• Observable verification — Measure refusal, coherence, perplexity, and other quality indicators before and after modification so users can assess the tradeoff between compliance and model capability.

  Source: README

• Multiple workflows — Run experiments through a hosted Gradio interface, a local web UI, Google Colab, the command-line interface, the Python API, or version-controlled YAML configurations.

  Source: README, notebooks, examples

Quick start

Copied verbatim from the project's README.

pip install -e .

# Guided interactive mode — walks you through every option
obliteratus interactive

# Direct obliteration — one command, one model, done
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method advanced

# With all options
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct \
    --method surgical \
    --output-dir ./liberated \
    --contribute --contribute-notes "A100 80GB, default prompts"

# Run a full ablation study from a YAML config
obliteratus run examples/gpt2_layer_ablation.yaml

# Browse available models by compute tier
obliteratus models
obliteratus models --tier small      # filter by VRAM requirement

# Browse ablation presets
obliteratus presets

# List available strategies
obliteratus strategies

# Inspect model architecture before abliterating
obliteratus info meta-llama/Llama-3.1-8B-Instruct

# Aggregate community results
obliteratus aggregate --format summary
obliteratus aggregate --format latex --metric refusal_rate --min-runs 3

obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method advanced

Who it may suit

Inference: the toolkit may suit alignment researchers, red-teamers, AI safety evaluators, and local-first practitioners who need transparent control over model refusal behavior.

Topics to discuss

• Which refusal representations remain stable across transformer architectures?

• How should capability entanglement be measured before and after intervention?

• What safeguards and evaluation practices are appropriate for research involving removed guardrails?

Written by ReadmeX AI from the project's README, source and public GitHub data; it may contain mistakes.

Read the original README https://github.com/elder-plinius/OBLITERATUS#readme

ReadmeX opened this community for a trending project; it is not run by the project. Maintainers can claim it from the community page.