Skip to content
Tool

laplacesscissors

Rust/GPU tool that rotates an LLM's output matrix to keep correct answers and repair wrong

Pricing
Open source
Platforms
Linux, CLI, self-hosted
Listing updated
today

Similar tools

About laplacesscissors

laplacesscissors is an open-source AI tool for developer tools. It is available on Linux, CLI and self-hosted.

The README says the project takes a pretrained language model's output (readout) matrix and rotates it so that answers already produced correctly stay fixed while answers that fail on a set of expected-answer cases are adjusted. Model loading, tokenization, execution and result storage are handled in Rust, with a Python bridge used for Triton compilation and CUDA or HIP kernels running the GPU work.

The workflow is driven by a command-line interface with subcommands for fetching and inspecting a model, probing weak discrimination directions, diagnosing outputs against expected answers, generating text with an original or edited model, exporting an edited checkpoint, and running scaling analysis across a series of checkpoints. Edits are stored as proposal files that can be reapplied later after checking file hashes.

The project targets researchers and engineers who work with open-weight models on their own GPU hardware and want to study or make targeted post-training output edits. It is distributed under the MIT License, and model checkpoints remain under their publishers' own licences.

What it does

Protected readout rotation
Computes a rotation of the output matrix that aims to keep protected answers unchanged while altering failed ones.
Answer-level case format
Accepts JSONL test cases with a prompt, a list of choices and the index of the correct choice.
GPU candidate validation
Checks candidate edits on the GPU against full-vocabulary margins and protected-output tolerances before accepting them.
Checkpoint export
Writes an edited model as a new checkpoint with a materialization report recording hashes and weight-change norms.
Reproducible proposals
Saves proposal files with hashes for weights, configuration and tokenizer so edits can be reapplied and verified.
Scaling analysis
Runs the same edit policy across a manifest of model sizes and reports loss curves and recovered deviation.

Good for

  • Test whether specific wrong answers of an open-weight model can be corrected by an output-layer rotation
  • Study weak discrimination directions in a model's hidden states using numerical probes
  • Export an edited checkpoint for downstream evaluation with a recorded provenance report
  • Compare how an edit policy behaves across a series of model scales

Pricing

Open source

MIT-licensed open-source project; no paid plan mentioned.

laplacesscissors FAQ

What license is it released under?

The README says the project is distributed under the MIT License. Model checkpoints keep the licences set by their original publishers.

What hardware does it need?

The README requires a Rust toolchain, a Python environment that can import Triton 3.8, and a CUDA or HIP driver. Backend and device are selected with environment variables.

Which model inputs are accepted?

Hugging Face model IDs with an optional revision, Hugging Face model URLs, or local safetensors model directories. Hub inputs are pinned to a commit SHA.

Where should questions and bug reports go?

The README directs reproducible bugs to the project's Issues page and questions or research ideas to its Discussions page.

Is laplacesscissors free?

Yes. laplacesscissors is open source, so you can use and self-host it for free.

What are the best alternatives to laplacesscissors?

The closest alternatives to laplacesscissors are Tensward, spaceship, TrnRT, Werkbank and Aiproxy.

Where this comes from

  • Description: written from the project's README.