Finch 7B Merge

A SLERP merge of two powerful 7B language models

Description

Finch is a 7B language model created by merging macadeliccc/WestLake-7B-v2-laser-truthy-dpo and SanjiWatsuki/Kunoichi-DPO-v2-7B using the SLERP method.

Quantized Models

Quantized versions of Finch are available:

• 6bpw EXL2 Quant
• GGUF Quants

Recommended Settings

For best results, use the ChatML format with the following sampler settings:

Temperature: 1.2 Min P: 0.2 Smoothing Factor: 0.2

Mergekit Configuration

base_model: macadeliccc/WestLake-7B-v2-laser-truthy-dpo dtype: float16 merge_method: slerp parameters: t: - filter: self_attn value: [0.0, 0.5, 0.3, 0.7, 1.0] - filter: mlp value: [1.0, 0.5, 0.7, 0.3, 0.0] - value: 0.5 slices: - sources: - layer_range: [0, 32] model: macadeliccc/WestLake-7B-v2-laser-truthy-dpo - layer_range: [0, 32] model: SanjiWatsuki/Kunoichi-DPO-v2-7B

Evaluation Results

Finch's performance on the Open LLM Leaderboard:

Metric	Value
Avg.	73.78
AI2 Reasoning Challenge (25-Shot)	71.59
HellaSwag (10-Shot)	87.87
MMLU (5-Shot)	64.81
TruthfulQA (0-shot)	67.96
Winogrande (5-shot)	84.14
GSM8k (5-shot)	66.34

Detailed results: https://huggingface.co/datasets/open-llm-leaderboard/details_antiven0m__finch