base_model:
- mistralai/Mistral-7B-v0.1
- berkeley-nest/Starling-LM-7B-alpha
- mlabonne/AlphaMonarch-7B
- cognitivecomputations/WestLake-7B-v2-laser
- senseable/garten2-7b
library_name: transformers
tags:
- mergekit
- merge
license: cc-by-nc-4.0
model-index:
- name: Starling_Monarch_Westlake_Garten-7B-v0.1
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: EQ-Bench
type: eq-bench
config: EQ-Bench
split: v2.1
args:
num_few_shot: 3
metrics:
- type: acc_norm
value: 80.01
name: self-reported
source:
url: https://github.com/EQ-bench/EQ-Bench
name: EQ-Bench v2.1
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 71.76
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 88.15
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 65.07
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 67.92
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 82.16
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 71.95
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1
name: Open LLM Leaderboard
Starling_Monarch_Westlake_Garten-7B-v0.1
After experimenting with density for a previous merge (containing similar models), I decided to experiment with weight gradients. My thought that was that if the merge was done with care and attention, I'd be able to create something greater than the sum of its parts. Hoping that, through a merge of really good models, I'd be able to create something greater than the sum of its parts.
I came across the EQ-Bench Benchmark (Paper) as part of my earlier testing. It is a very light and quick benchmark that yields powerful insights into how well the model performs in emotional intelligence related prompts. As part of this process, I tried to figure out if there was a way to determine an optimal set of gradient weights that would lead to the most successful merge as measured against EQ-Bench. At first, my goal was to simply exceed WestLake-7B, but then I kept pushing to see what I could come up with. Too late in the process, I learned that dare_ties has a random element to it. Valuable information for next time, I guess. After concluding that project, I began collecting more data, this time setting a specified seed in mergekit for reproducibility. As I was collecting data, I hit the goal I had set for myself. This model is not a result of the above work but is the genesis of how this model came to be.
I present, Starling_Monarch_Westlake_Garten-7B-v0.1, the only 7B model to score > 80 on the EQ-Bench v2.1 benchmark found here, outscoring larger models like abacusai/Smaug-72B-v0.1 and cognitivecomputations/dolphin-2.2-70b
It also surpasses its components in the GSM8K benchmark, with a score of 71.95. I'll be looking to bring out more logic and emotion in the next evolution of this model.
It also earned 8.109 on MT-Bench(paper), outscoring Chat-GPT 3.5 and Claude v1.
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the DARE TIES merge method using mistralai/Mistral-7B-v0.1 as a base. The seed for this merge is 176
Models Merged
The following models were included in the merge:
- berkeley-nest/Starling-LM-7B-alpha
- mlabonne/AlphaMonarch-7B
- cognitivecomputations/WestLake-7B-v2-laser
- senseable/garten2-7b
Configuration
The following YAML configuration was used to produce this model:
models:
- model: mistralai/Mistral-7B-v0.1
# No parameters necessary for base model
- model: cognitivecomputations/WestLake-7B-v2-laser
parameters:
density: 0.58
weight: [0.3877, 0.1636, 0.186, 0.0502]
- model: senseable/garten2-7b
parameters:
density: 0.58
weight: [0.234, 0.2423, 0.2148, 0.2775]
- model: berkeley-nest/Starling-LM-7B-alpha
parameters:
density: 0.58
weight: [0.1593, 0.1573, 0.1693, 0.3413]
- model: mlabonne/AlphaMonarch-7B
parameters:
density: 0.58
weight: [0.219, 0.4368, 0.4299, 0.331]
merge_method: dare_ties
base_model: mistralai/Mistral-7B-v0.1
parameters:
int8_mask: true
dtype: bfloat16
Table of Benchmarks
Open LLM Leaderboard
| Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K | |
|---|---|---|---|---|---|---|---|
| giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1 | 74.9 | 71.76 | 88.15 | 65.07 | 67.92 | 84.53 | 71.95 |
| mlabonne/AlphaMonarch-7B | 75.99 | 73.04 | 89.18 | 64.4 | 77.91 | 84.69 | 66.72 |
| senseable/WestLake-7B-v2 | 74.68 | 73.04 | 88.65 | 64.71 | 67.06 | 86.98 | 67.63 |
| berkeley-nest/Starling-LM-7B-alpha | 67.13 | 63.82 | 84.9 | 63.64 | 46.39 | 80.58 | 62.4 |
| senseable/garten2-7b | 72.65 | 69.37 | 87.54 | 65.44 | 59.5 | 84.69 | 69.37 |
Yet Another LLM Leaderboard benchmarks
| Model | AGIEval | GPT4All | TruthfulQA | Bigbench | Average |
|---|---|---|---|---|---|
| giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1 | 44.99 | 76.93 | 68.04 | 47.71 | 59.42 |
| mlabonne/AlphaMonarch-7B | 45.37 | 77 | 78.39 | 50.2 | 62.74 |
| berkeley-nest/Starling-LM-7B-alpha | 42.06 | 72.72 | 47.33 | 42.53 | 51.16 |
Misc. Benchmarks
| MT-Bench | EQ-Bench v2.1 | |
|---|---|---|
| giraffe176/Starling_Monarch_Westlake_Garten-7B-v0.1 | 8.109375 | 80.01 (3 Shot, ChatML, ooba) |
| mlabonne/AlphaMonarch-7B | 8.23750 | 76.08 |
| senseable/WestLake-7B-v2 | X | 78.7 |
| berkeley-nest/Starling-LM-7B-alpha | 8.09 | 68.69 (1 Shot, ChatML, ooba) |
| senseable/garten2-7b | X | 75.03 |
| claude-v1 | 7.900000 | 76.83 |
| gpt-3.5-turbo | 7.943750 | 71.74 |
| (Paper) | (Paper) Leaderboard |