Lewdiculous's picture
Update README.md
797003d verified
metadata
library_name: transformers
tags:
  - mergekit
  - merge
  - roleplay
  - mistral
inference: false
license: apache-2.0

This repository hosts GGUF-IQ-Imatrix quants for vicgalle/RoleBeagle-11B.

What does "Imatrix" mean?

It stands for Importance Matrix, a technique used to improve the quality of quantized models. The Imatrix is calculated based on calibration data, and it helps determine the importance of different model activations during the quantization process. The idea is to preserve the most important information during quantization, which can help reduce the loss of model performance, especially when the calibration data is diverse. [1] [2]

For imatrix data generation, kalomaze's groups_merged.txt with added roleplay chats was used, you can find it here. This was just to add a bit more diversity to the data.

Steps:

Base⇢ GGUF(F16)⇢ Imatrix-Data(F16)⇢ GGUF(Imatrix-Quants)

Using the latest llama.cpp at the time.

    quantization_options = [
        "Q4_K_M", "Q4_K_S", "IQ4_XS", "Q5_K_M", "Q5_K_S",
        "Q6_K", "Q8_0", "IQ3_M", "IQ3_S", "IQ3_XXS"
    ]

Original model information:

RoleBeagle-11B

img

A DPO-finetune from vicgalle/CarbonBeagle-11B-truthy over a subset of OpenHermesPreferences containting RP conversations. It keeps most of the intelligence from CarbonBeagle-11B, and hopefuly can role-play better.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 76.06
AI2 Reasoning Challenge (25-Shot) 72.35
HellaSwag (10-Shot) 89.77
MMLU (5-Shot) 66.35
TruthfulQA (0-shot) 77.92
Winogrande (5-shot) 84.06
GSM8k (5-shot) 65.88