laserxtral-exl2

Exllama v2 Quantizations of laserxtral

Using <a href="https://github.com/turboderp/exllamav2/releases/tag/v0.0.11">turboderp's ExLlamaV2 v0.0.11</a> for quantization.

The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)

Discord: https://discord.gg/cognitivecomputations

Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.

Conversion was done using the default calibration dataset.

Default arguments used.

Original model: https://huggingface.co/cognitivecomputations/laserxtral

<a href="https://huggingface.co/cognitivecomputations/laserxtral-exl2/tree/6.5">6.5 bits per weight</a>

<a href="https://huggingface.co/cognitivecomputations/laserxtral-exl2/tree/4">4 bits per weight</a>

<a href="https://huggingface.co/cognitivecomputations/laserxtral-exl2/tree/3">3 bits per weight</a>

<a href="https://huggingface.co/cognitivecomputations/laserxtral-exl2/tree/2">2 bits per weight</a>

Credit to Bartowski for help and model card formatting

Original Model Card Below

by David, Fernando and Eric

Sponsored by: VAGO Solutions

Join our Discord! https://discord.gg/vT3sktQ3zb

An experimentation regarding 'lasering' each expert to denoise and enhance model capabilities.

This model has half size in comparison to the Mixtral 8x7b Instruct. And it basically has the same level of performance (we are working to get a better MMLU score).

Laserxtral - 4x7b (all, except for base, lasered using laserRMT)

This model is a Mixture of Experts (MoE) made with mergekit (mixtral branch). It uses the following base models:

• cognitivecomputations/dolphin-2.6-mistral-7b-dpo
• mlabonne/Marcoro14-7B-slerp (base)
• beowolx/CodeNinja-1.0-OpenChat-7B
• Q-bert/MetaMath-Cybertron-Starling
• WizardLM/WizardMath-7B-V1.1

It follows the implementation of laserRMT @ https://github.com/cognitivecomputations/laserRMT

Here, we are controlling layers checking which ones have lower signal to noise ratios (which are more subject to noise), to apply Laser interventions, still using Machenko Pastur to calculate this ratio.

We intend to be the first of a family of experimentations being carried out @ Cognitive Computations.

In this experiment we have observed very high truthfulness and high reasoning capabilities.