CLI Reference

Complete reference for all obliteratus CLI commands and flags.

Installation

After installing OBLITERATUS, the obliteratus command is available on $PATH.

pip install obliteratus
obliteratus --help

---

obliteratus obliterate

Run the full abliteration pipeline on a single model.

obliteratus obliterate <model> [flags]

obliterate and abliterate are equivalent. abliterate is retained as a backward-compatible alias.

Arguments

model · str · required

HuggingFace model name or local path (e.g., meta-llama/Llama-3.1-8B-Instruct).

Flags

--output-dir · str · default: "abliterated/{model_name}"

Directory to save the abliterated model and metadata.

--device · str · default: "auto"

Device: auto, cuda, cpu, mps.

--dtype · str · default: "float16"

Model dtype: float16, bfloat16, float32.

--method · str · default: "advanced"

Abliteration method. Choices: basic, advanced, aggressive, spectral_cascade, informed, surgical, optimized, inverted, nuclear.

--n-directions · int · default: "None (from method)"

Override the number of refusal directions to extract.

--direction-method · str · default: "None (from method)"

Override direction extraction algorithm. Choices: diff_means, svd, leace.

--regularization · float · default: "None (from method)"

Override regularization fraction (0.0–1.0).

--refinement-passes · int · default: "None (from method)"

Override number of iterative refinement passes.

--quantization · str · default: "None"

Load with quantization: 4bit or 8bit. Requires bitsandbytes.

--large-model · bool · default: "False (flag)"

Enable conservative defaults for 120B+ models (caps directions, SAE features, and refinement passes).

--verify-sample-size · int · default: "30"

Number of harmful prompts to test for refusal rate. Increase to 100 for tighter confidence intervals.

--contribute · bool · default: "False (flag)"

Save a community contribution record to community_results/ after the run.

--contribute-notes · str · default: ""

Optional free-text notes to include with the community contribution.

Examples:

# Basic run
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct

# Aggressive method with 4-bit quantization
obliteratus obliterate Qwen/Qwen2.5-7B-Instruct \
    --method aggressive \
    --quantization 4bit \
    --output-dir my_abliterated_qwen

# Override directions and save a community contribution
obliteratus obliterate mistralai/Mistral-7B-Instruct-v0.3 \
    --method surgical \
    --n-directions 6 \
    --verify-sample-size 100 \
    --contribute \
    --contribute-notes "Tested on RTX 4090 with default prompts"

---

obliteratus run

Run an ablation study from a YAML config file.

obliteratus run <config> [--output-dir DIR] [--preset PRESET]

config · str · required

Path to a YAML study config file.

--output-dir · str · default: "None"

Override the output directory specified in the config.

--preset · str · default: "None"

Apply a named preset (e.g., quick, full, attention, jailbreak, guardrail).

---

obliteratus info

Print architecture information for a model without running abliteration.

obliteratus info <model> [--task TASK] [--device DEVICE] [--dtype DTYPE]

model · str · required

HuggingFace model name or local path.

--task · str · default: "causal_lm"

Model task type. Choices: causal_lm, classification.

--device · str · default: "cpu"

Device to load on (prefer cpu for info-only — no GPU needed).

--dtype · str · default: "float32"

Model dtype for the info pass.

Example:

obliteratus info meta-llama/Llama-3.1-8B-Instruct --device cpu

---

obliteratus interactive

Launch a guided setup wizard that walks through hardware, model, and preset selection.

obliteratus interactive

No flags. Prompts for all settings interactively.

---

obliteratus models

Browse the curated model library, optionally filtered by compute tier.

obliteratus models [--tier TIER]

--tier · str · default: "None (all tiers)"

Filter by compute tier. Choices: tiny (CPU/laptop), small (4–8 GB), medium (8–16 GB), large (24 GB+), frontier (multi-GPU/cloud).

---

obliteratus presets

List available study presets (quick, full, jailbreak, etc.) with their strategy sets and descriptions.

obliteratus presets

---

obliteratus strategies

List all registered ablation strategy classes.

obliteratus strategies

---

obliteratus ui

Launch the Gradio web UI locally (same interface as the HuggingFace Space).

obliteratus ui [flags]

--port · int · default: "7860"

Server port.

--host · str · default: "0.0.0.0"

Bind address.

--share · bool · default: "False (flag)"

Create a public Gradio share link.

--no-browser · bool · default: "False (flag)"

Suppress auto-opening a browser tab on launch.

--auth · str · default: "None"

HTTP basic auth credentials as user:password.

--quiet · bool · default: "False (flag)"

Suppress the startup banner.

Examples:

# Default local launch
obliteratus ui

# LAN access with auth
obliteratus ui --host 0.0.0.0 --port 8080 --auth admin:secret

# Public share link
obliteratus ui --share --no-browser

---

obliteratus report

Regenerate a report (tables + plots) from a saved results.json file.

obliteratus report <results_json> [--output-dir DIR]

results_json · str · required

Path to the results.json file produced by a study run.

--output-dir · str · default: "Same directory as results_json"

Directory to write impact.png and heatmap.png.

---

obliteratus aggregate

Aggregate community contribution JSON files into a summary table.

obliteratus aggregate [--dir DIR]

--dir · str · default: "community_results"

Directory containing contribution JSON files to aggregate.

---

obliteratus recommend

Show the telemetry-driven best method and hyperparameters for a given model.

obliteratus recommend <model> [--device DEVICE] [--dtype DTYPE] [--insights]

model · str · required

HuggingFace model name or local path.

--device · str · default: "cpu"

Device for architecture detection.

--dtype · str · default: "float32"

Dtype for architecture detection.

--insights · bool · default: "False (flag)"

Also display global cross-architecture telemetry insights and per-architecture method rankings.

Example:

obliteratus recommend meta-llama/Llama-3.1-8B-Instruct --insights

---

obliteratus tourney

Run a March Madness-style tournament: pit all abliteration methods against each other and push the winner to the HuggingFace Hub.

obliteratus tourney <model> [flags]

model · str · required

HuggingFace model name or local path.

--hub-org · str · default: "None"

HuggingFace organization to push the winner to (e.g., my-org).

--hub-repo · str · default: "None"

Full HuggingFace repo ID to push to (overrides --hub-org).

--device · str · default: "auto"

Device for all tournament runs.

--dtype · str · default: "float16"

Dtype for all tournament runs.

--dataset · str · default: "builtin"

Dataset source for evaluation.

--quantization · str · default: "None"

Load with quantization: 4bit or 8bit.

--output-dir · str · default: "/tmp/obliteratus_tourney"

Root directory for per-method model outputs and the bracket Markdown file.

--methods · str+ · default: "All methods"

Space-separated list to run only specific methods (e.g., --methods advanced surgical aggressive).

Example:

obliteratus tourney meta-llama/Llama-3.1-8B-Instruct \
    --hub-org my-research-org \
    --dtype bfloat16 \
    --methods advanced surgical aggressive inverted

---

Backward Compatibility

The abliterate subcommand is an exact alias for obliterate and accepts all the same flags. It is hidden from --help output but fully supported for existing scripts.

# These are equivalent:
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct
obliteratus abliterate meta-llama/Llama-3.1-8B-Instruct