File size: 3,485 Bytes
f38676f f1e2cad f38676f f1e2cad f38676f 1733fd5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 |
---
library_name: transformers
tags:
- mergekit
- merge
license: llama3.1
---
# Llama-3.1-SuperNova-Lite_TIES_with_Base
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).
## Merge Details/Method
This is a merge of [arcee-ai/Llama-3.1-SuperNova-Lite](https://huggingface.co/arcee-ai/Llama-3.1-SuperNova-Lite) with its base [meta-llama/Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B) (the base model being: the model which the instruct model was fine-tuned on - even though in our case, [arcee-ai/Llama-3.1-SuperNova-Lite](https://huggingface.co/arcee-ai/Llama-3.1-SuperNova-Lite), was fine-tuned, etc on top of [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) and not directly on top of [meta-llama/Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B))
This model was merged using the [TIES](https://arxiv.org/abs/2306.01708) merge method using meta-llama/Llama-3.1-8B as a base.
The merge was inspired by RomboDawg's ([Replete-AI](https://huggingface.co/Replete-AI)) TIES merge of [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) with its base [Qwen/Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B), which topped the OpenLLM Learderboard for the highest Average score for a 7B parameter model.
After experimenting and discussing/researching the merge with Rombodawg, I looked into mergekit's TIES merge method some more, which led me to find a pertinent parameter that we weren't utilizing for our TIES merge: density. I decided to use density along with the weight parameter to see if we could restore some of the instruction following that our merges seemed to lack in comparison to the original Instruct model. The resulant merges turned out to be great! By using the density parameter along with the weight parameter, we were able to restore more of the Instruction following which was diminished and/or not present when solely using the weight parameter for our TIES merge.
The way this works is: the Instruct model is TIES merged with the base model, with the weight = 1 and density = 1. After the merge is complete, the merge's .json config files (excluding 'model.safetensors.index.json') are replaced with the original Instruct's .json config files.
### Models Merged
The following models were included in the merge:
* /Users/jsarnecki/opt/Workspace/arcee-ai/Llama-3.1-SuperNova-Lite
### Configuration
The following YAML configuration was used to produce this model:
```yaml
models:
- model: "/Users/jsarnecki/opt/Workspace/arcee-ai/Llama-3.1-SuperNova-Lite"
parameters:
weight: 1
density: 1
- model: "/Users/jsarnecki/opt/Workspace/arcee-ai/Llama-3.1-SuperNova-Lite"
parameters:
weight: 1
density: 1
merge_method: ties
base_model: "/Users/jsarnecki/opt/Workspace/meta-llama/Llama-3.1-8B"
parameters:
density: 1
normalize: true
int8_mask: true
dtype: bfloat16
```
Open LLM Leaderboard Evaluation Results
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/results/blob/main/Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base/results_2024-10-03T03-40-13.800772.json)
| Metric |Value|
|-------------------|----:|
|Average Score |43.07|
|IFEval (0-Shot) |80.96|
|BBH (3-Shot) |51.10|
|MATH Lvl 5 (4-Shot)|15.56|
|GPQA (0-shot) |30.96|
|MuSR (0-shot) |41.01|
|MMLU-PRO (5-shot) |38.80|
|