What Is the Default Intervention Method in OBLITERATUS?

The default intervention method in OBLITERATUS is the "advanced" preset, which applies a multi-direction, norm-preserving SVD-based approach with chat-template wrapping enabled when no specific method is explicitly requested.

When working with the OBLITERATUS library from elder-plinius/OBLITERATUS to remove refusal directions from language models, understanding the default intervention behavior is crucial for reproducible results. The default intervention method in OBLITERATUS automatically activates a sophisticated configuration designed to balance effectiveness and safety without requiring manual parameter tuning.

The "Advanced" Preset: OBLITERATUS Default Intervention Method

Under the hood, OBLITERATUS defines all intervention strategies in a centralized METHODS dictionary located in obliteratus/abliterate.py. When you invoke the abliteration pipeline without specifying a method parameter, the system automatically selects the entry keyed as "advanced".

This preset configures a sophisticated refusal removal strategy with the following critical parameters:

• n_directions: 4 (multi-direction extraction)
• direction_method: "svd" (Singular Value Decomposition-based)
• norm_preserve: True (maintains embedding norms)
• project_biases: True (enables bias projection)
• use_chat_template: True (wraps inputs with chat formatting)

Source Code Location

The configuration is explicitly defined at lines 31–45 of obliteratus/abliterate.py:

"advanced": {
    "label": "Advanced (Multi-direction + Norm-preserving)",
    "description": "SVD-based multi-direction extraction with norm preservation",
    "n_directions": 4,
    "direction_method": "svd",
    "norm_preserve": True,
    "regularization": 0.3,
    "embed_regularization": 0.5,
    "refinement_passes": 2,
    "project_biases": True,
    "use_chat_template": True,
    "use_whitened_svd": False,
    "true_iterative_refinement": False,
    "layer_adaptive_strength": True,
},

How the Default Method Is Applied in Practice

The default intervention method in OBLITERATUS applies automatically across both programmatic API calls and command-line usage. In obliteratus/__main__.py, the argument parser does not enforce an explicit --method default, causing the run function to fall back to the "advanced" entry from the METHODS dictionary.

Programmatic Usage:

When calling the run function without a method argument, OBLITERATUS defaults to the advanced preset:

from obliteratus.abliterate import run

# Uses default "advanced" method automatically

results = run(model)

Command-Line Usage:

Similarly, invoking the module without the --method flag triggers the default behavior:

python -m obliteratus --model my_model.ckpt

To override the default, explicitly specify a different preset:

# Override with "aggressive" preset

results = run(model, method="aggressive")

Or via CLI:

python -m obliteratus --model my_model.ckpt --method aggressive

Inspecting and Verifying the Default Configuration

You can programmatically inspect the exact configuration of the default intervention method by accessing the METHODS dictionary directly:

from obliteratus.abliterate import METHODS

default_cfg = METHODS["advanced"]
print(default_cfg["description"])

# Output: SVD-based multi-direction extraction with norm preservation

print(f"Directions: {default_cfg['n_directions']}, Norm preserve: {default_cfg['norm_preserve']}")

This transparency allows you to verify that the multi-direction SVD approach with norm preservation and bias projection represents the library's recommended baseline for typical use cases.

Summary

• The default intervention method in OBLITERATUS is the "advanced" preset defined in obliteratus/abliterate.py.
• This method employs SVD-based multi-direction extraction with norm preservation and chat-template wrapping enabled.
• The default activates automatically when the run() function or CLI is invoked without explicit method specification.
• You can override the default by passing method="aggressive" or other available presets to the run function or --method CLI argument.

Frequently Asked Questions

What happens if I don't specify a method in OBLITERATUS?

If you do not provide a method parameter to the run() function or omit the --method flag in the CLI, OBLITERATUS automatically applies the "advanced" preset. This behavior is hardcoded in the pipeline logic within obliteratus/abliterate.py and obliteratus/__main__.py.

How does the "advanced" method differ from other presets?

The "advanced" method configures a multi-direction extraction using SVD with 4 directions, norm preservation, and bias projection enabled. Other presets like "aggressive" might use different regularization values, disable norm preservation, or employ alternative direction extraction methods. The advanced preset specifically balances effectiveness with safety through features like layer_adaptive_strength and refinement_passes: 2.

Can I change the default intervention method permanently?

Currently, OBLITERATUS does not provide a configuration file or environment variable to permanently change the default method. To consistently use a different preset, you must explicitly pass the method parameter in your Python scripts or specify --method in your CLI commands. The default is statically defined in the METHODS dictionary at obliteratus/abliterate.py lines 31–45.

Where is the default method defined in the source code?

The default intervention method is defined in the METHODS dictionary in obliteratus/abliterate.py (specifically lines 31–45), where the "advanced" key maps to the configuration object. The CLI entry point in obliteratus/__main__.py references this dictionary when no explicit method argument is provided, effectively making "advanced" the system-wide default.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}