ZeroXClem commited on
Commit
ad8b3a6
β€’
1 Parent(s): 5d53dfd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +106 -18
README.md CHANGED
@@ -2,43 +2,82 @@
2
  license: apache-2.0
3
  tags:
4
  - merge
5
- - mergekit
6
- - lazymergekit
 
 
 
 
 
 
 
 
7
  - bunnycore/Best-Mix-Llama-3.1-8B
8
- - USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B
9
- - Weyaxi/Einstein-v6.1-Llama3-8B
10
  ---
11
 
12
- # ZeroXClem/Llama3.1-BestMix-Chem-Einstein-8B
13
 
14
- ZeroXClem/Llama3.1-BestMix-Chem-Einstein-8B is a merge of the following models using [mergekit](https://github.com/cg123/mergekit):
15
- * [bunnycore/Best-Mix-Llama-3.1-8B](https://huggingface.co/bunnycore/Best-Mix-Llama-3.1-8B)
16
- * [USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B](https://huggingface.co/USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B)
17
- * [Weyaxi/Einstein-v6.1-Llama3-8B](https://huggingface.co/Weyaxi/Einstein-v6.1-Llama3-8B)
18
 
19
- ## 🧩 Configuration
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20
 
21
  ```yaml
 
 
22
  models:
23
  - model: bunnycore/Best-Mix-Llama-3.1-8B
24
  parameters:
25
- density: [1, 0.7, 0.5] # density gradient, focusing on overall balance
26
  weight: 1.0
27
 
28
  - model: USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B
29
  parameters:
30
- density: 0.6 # chemistry focus with significant contribution
31
- weight: [0.3, 0.7, 1.0] # weight gradient, increasing focus on chemistry tasks
32
 
33
  - model: Weyaxi/Einstein-v6.1-Llama3-8B
34
  parameters:
35
- density: 0.4 # focus on conversational adaptability and long-form generation
36
  weight:
37
  - filter: mlp
38
- value: 0.5 # stronger contribution to MLP tasks
39
  - filter: self_attn
40
- value: 0.7 # slightly higher weight on attention mechanisms
41
- - value: 0.5 # balanced contribution for general tasks
42
 
43
  merge_method: ties
44
  base_model: bunnycore/Best-Mix-Llama-3.1-8B
@@ -47,4 +86,53 @@ parameters:
47
  int8_mask: true
48
  dtype: float16
49
 
50
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  license: apache-2.0
3
  tags:
4
  - merge
5
+ - TIES
6
+ - Llama3
7
+ - BestMix
8
+ - Chemistry
9
+ - Einstein
10
+ - instruction-following
11
+ - conversational
12
+ - long-form-generation
13
+ - scientific
14
+ base_model:
15
  - bunnycore/Best-Mix-Llama-3.1-8B
 
 
16
  ---
17
 
 
18
 
19
+ # **ZeroXClem/Llama3.1-BestMix-Chem-Einstein-8B**
 
 
 
20
 
21
+ **Llama3.1-BestMix-Chem-Einstein-8B** is an innovative, meticulously blended model designed to excel in **instruction-following**, **chemistry-focused tasks**, and **long-form conversational generation**. This model fuses the **best qualities** of multiple Llama3-based architectures, making it highly versatile for both general and specialized tasks. πŸ’»πŸ§ βœ¨
22
+
23
+ ## 🌟 **Family Tree**
24
+
25
+ This model is the result of merging the following:
26
+
27
+ - [**bunnycore/Best-Mix-Llama-3.1-8B**](https://huggingface.co/bunnycore/Best-Mix-Llama-3.1-8B): A balanced blend of top Llama models, optimized for general performance across reasoning, instruction-following, and math.
28
+ - [**USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B**](https://huggingface.co/USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B): A model specialized in **scientific knowledge** and **chemistry**, excelling in chemistry benchmarks.
29
+ - [**Weyaxi/Einstein-v6.1-Llama3-8B**](https://huggingface.co/Weyaxi/Einstein-v6.1-Llama3-8B): Fine-tuned for **long-form generation**, **conversation-heavy tasks**, and optimized with cutting-edge techniques for efficient memory usage and fast performance.
30
+
31
+ ---
32
+
33
+ ## 🧬 **Model Lineage**
34
+
35
+ ### **A: bunnycore/Best-Mix-Llama-3.1-8B**
36
+
37
+ - A masterful **blend** of several Llama3 models like **Aurora_faustus**, **TitanFusion**, and **OpenMath2**.
38
+ - Provides a **balanced performance** in a variety of tasks such as reasoning, math, and instruction-following.
39
+ - Key contributor to the **overall versatility** of the merged model.
40
+
41
+ ### **B: USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B**
42
+
43
+ - Specializes in **chemistry** and **scientific knowledge**, outperforming many larger models in **chemistry benchmarks**.
44
+ - Adds **scientific rigor** and domain-specific expertise to the merged model, making it perfect for scientific and academic tasks.
45
+
46
+ ### **C: Weyaxi/Einstein-v6.1-Llama3-8B**
47
+
48
+ - Fine-tuned on a wide range of **instructive** and **conversational datasets** like **WizardLM**, **Alpaca**, and **ShareGPT**.
49
+ - Optimized for **long-form text generation** and enhanced with **xformers attention** and **flash attention** techniques for better performance.
50
+ - Key player in **dialogue-based tasks** and **long conversation generation**.
51
+
52
+ ---
53
+
54
+ ## πŸ› οΈ **Merge Details**
55
+
56
+ This model was merged using the **TIES merge method**, ensuring a smooth integration of the key strengths from each contributing model. Here's the configuration used:
57
 
58
  ```yaml
59
+ yaml
60
+ Copy code
61
  models:
62
  - model: bunnycore/Best-Mix-Llama-3.1-8B
63
  parameters:
64
+ density: [1, 0.7, 0.5]
65
  weight: 1.0
66
 
67
  - model: USTC-KnowledgeComputingLab/Llama3-KALE-LM-Chem-1.5-8B
68
  parameters:
69
+ density: 0.6
70
+ weight: [0.3, 0.7, 1.0]
71
 
72
  - model: Weyaxi/Einstein-v6.1-Llama3-8B
73
  parameters:
74
+ density: 0.4
75
  weight:
76
  - filter: mlp
77
+ value: 0.5
78
  - filter: self_attn
79
+ value: 0.7
80
+ - value: 0.5
81
 
82
  merge_method: ties
83
  base_model: bunnycore/Best-Mix-Llama-3.1-8B
 
86
  int8_mask: true
87
  dtype: float16
88
 
89
+ ```
90
+
91
+ ---
92
+
93
+ ## 🎯 **Key Features & Capabilities**
94
+
95
+ ### **1. Instruction Following & General Reasoning**:
96
+
97
+ With the foundation of **Best-Mix**, this model excels in **general-purpose reasoning**, instruction-following, and tasks that require high adaptability.
98
+
99
+ ### **2. Scientific & Chemistry Expertise**:
100
+
101
+ Thanks to the contribution from **KALE-LM-Chem**, this model shines in **scientific research**, particularly **chemistry-focused tasks**, making it ideal for academic and research purposes.
102
+
103
+ ### **3. Long-Form & Conversational Mastery**:
104
+
105
+ With **Einstein-v6.1**, the model handles **long-form generation** effortlessly, excelling in extended conversations and structured dialogue applications.
106
+
107
+ ---
108
+
109
+ ## πŸš€ **Performance Benchmarks**
110
+
111
+ While still in its early stages, **Llama3.1-BestMix-Chem-Einstein-8B** is expected to perform well across a variety of benchmarks, including:
112
+
113
+ - **Chemistry-focused benchmarks** (KALE-LM-Chem)
114
+ - **Instruction-following tasks** (Best-Mix)
115
+ - **Conversational AI** and **long-form text generation** (Einstein-v6.1)
116
+
117
+ Further testing and evaluation will continue to refine this model's capabilities.
118
+
119
+ ---
120
+
121
+ ## πŸ“œ **License**
122
+
123
+ This model is open-sourced under the **Apache-2.0 License**, allowing free use and modification with proper attribution.
124
+
125
+ ---
126
+
127
+ ## πŸ’‘ **Tags**
128
+
129
+ - `merge`
130
+ - `TIES`
131
+ - `BestMix`
132
+ - `Chemistry`
133
+ - `Einstein`
134
+ - `instruction-following`
135
+ - `long-form-generation`
136
+ - `conversational`
137
+
138
+ ---