SHOWCASE ๐
Collection
Models you at least try
โข
2 items
โข
Updated
โข
1
This is a merge of pre-trained language models created using mergekit.
This model was merged using the DARE TIES merge method using unsloth/Meta-Llama-3.1-8B-Instruct as a base.
The following models were included in the merge:
The following YAML configuration was used to produce this model:
base_model: unsloth/Meta-Llama-3.1-8B-Instruct
dtype: bfloat16
merge_method: dare_ties
parameters:
int8_mask: 1.0
slices:
- sources:
- layer_range: [0, 32]
model: akjindal53244/Llama-3.1-Storm-8B
parameters:
density: 0.8
weight: 0.25
- layer_range: [0, 32]
model: arcee-ai/Llama-3.1-SuperNova-Lite
parameters:
density: 0.8
weight: 0.33
- layer_range: [0, 32]
model: Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2
parameters:
density: 0.8
weight: 0.42
- layer_range: [0, 32]
model: unsloth/Meta-Llama-3.1-8B-Instruct
tokenizer_source: base
Detailed results can be found here! Based on the listed rankings as of 4/12/24, is the top-rank 8B model.
Metric | Value |
---|---|
Avg. | 30.07 |
IFEval (0-Shot) | 80.29 |
BBH (3-Shot) | 31.61 |
MATH Lvl 5 (4-Shot) | 21.15 |
GPQA (0-shot) | 6.94 |
MuSR (0-shot) | 8.24 |
MMLU-PRO (5-shot) | 32.18 |
Personal recommendations are to use a i1-Q4_K_M quant with these settings:
num_ctx = 4096
repeat_penalty = 1.2
temperature = 0.85
tfs_z = 1.4
top_k = 0 # Change to 40+ if you're roleplaying
top_p = 1 # Change to 0.9 if top_k > 0
Other recommendations can be found on this paper on mobile LLMs, this paper on balancing model parameters, and this Reddit post about tweaking Llama 3.1 parameters.