FAQ¶

Who makes SafeTune, and how does it relate to Lexsi Labs?¶

SafeTune is developed and maintained by Lexsi Labs, the AI research arm of Aurionpro — an NSE-listed enterprise technology company (USD 140M revenue) with operations in banking, transit, capital markets, healthcare, and legal verticals.

SafeTune is source-available under LSAL v1.1 and free to use independently for research, academic, and personal purposes. For commercial licensing, enterprise support, custom deployment, compliance consulting, or custom method development on your own models and data, contact Lexsi Labs through lexsi.ai.

The library is the direct output of a controlled research study on safety-after-fine-tuning across Aurionpro's customer deployments. Every method has been tested on real production checkpoints, not just academic benchmarks.

What is SafeTune?¶

A library of LLM-safety methods. Each task (harden, recover, steer, unlearn, interpret, evaluate) has many independent methods that solve it by different mechanisms. You pick one per task.

Is SafeTune a pipeline?¶

No. A pipeline implies sequence (diagnose → harden → recover → verify). SafeTune doesn't work that way. Each task has alternatives — you pick one, not chain them. RESTA and C-ΔΘ and LoX are all ways to recover a drifted model; they are alternatives, not stages.

Does SafeTune work on CPU?¶

The core library imports cleanly on CPU. The quickstart demos run on Qwen/Qwen2.5-0.5B-Instruct with no GPU. Training-based methods (harden, unlearn) benefit from a GPU but work on CPU for small models.

What models does SafeTune support?¶

Any Hugging Face transformers-compatible causal LM. The audit used Llama-3.2-3B-Instruct; the quickstarts use Qwen2.5-0.5B-Instruct. Larger models show sharper safety effects.

What's the difference between Recover and Unlearn?¶

Recover edits weights directly — no training, no optimizer steps. You provide a drifted model + base/aligned references.

Unlearn trains — it runs optimizer steps on a forget set + retain set to erase a specific capability while preserving everything else.

Can I chain methods?¶

You can call them in sequence, but you shouldn't unless you understand why. Each method is an alternative for the same task. The combined example in the guides is labelled "illustrative only" — it shows the API shape, not a recommended workflow.

How do I cite SafeTune?¶

@misc{seth2026safetune,
  title  = {SafeTune: A Unified Library for Preserving and Restoring
            Safety in Fine-Tuned {LLM}s},
  author = {Seth, Pratinav and Kaushal, Anshul and Sadhu, Saisab and
            Sankarapu, Vinay Kumar},
  year   = {2026},
  note   = {Pratinav Seth, Anshul Kaushal, and Saisab Sadhu contributed equally.},
}

A machine-readable CITATION.cff is also available at the repo root.

How do I report a faithfulness issue?¶

Open a faithfulness report. Include the method name, what the paper says, what the code does, and file:line evidence.