File size: 5,333 Bytes
7acff0e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c8a2ccc
5610867
 
 
 
c8a2ccc
47511be
699ec60
47511be
 
 
 
 
7acff0e
1c46ce6
7acff0e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
---
language:
- en
license: apache-2.0
---

<div align="center">
  <b style="font-size: 40px;">Zion_Alpha_Instruction_Tuned_SLERP</b>


</div>


<img src="https://i.imgur.com/e1LEQ18.png" alt="Zion_Alpha_Instruction_Tuned_SLERP" style="width: 50%; min-width: 400px; display: block; margin: auto;">



# Model Details

Zion_Alpha is the first **REAL** Hebrew model in the world. This version WAS fine tuned for tasks. I did the finetune using SOTA techniques and using my insights from years of underwater basket weaving. If you wanna offer me a job, just add me on Facebook.


# Another world record broken by Zion_Alpha!
On **June 10th, 2024**, this model achieved the **highest sentiment analysis score in the world** for Hebrew LLMs, with an impressive **70.3**, surpassing even a **35B** model that's five times its size!
<div align="center">
<img src="https://i.imgur.com/yg6CJoz.png" alt="Zion_Alpha SNLI Score" style="width: 80%; min-width: 700px; display: block; margin: auto;">
</div>


# Future Plans
My previous LLM, Zion_Alpha, set a world record on Hugging Face by achieving the highest SNLI score for Hebrew open LLMs at 84.05. The current model, a SLERP merge, achieved a lower SNLI score but still surprised everyone by securing the highest sentiment analysis score of 70.3. This demonstrates significant untapped potential in optimizing the training process, showing that 7B models can deliver far more performance in Hebrew than previously thought possible. This will be my last Hebrew model for a while, as I have other adventures to pursue.
# Looking for Sponsors
Since all my work is done on-premises, I am constrained by my current hardware. I would greatly appreciate any support in acquiring an A6000, which would enable me to train significantly larger models much faster.

# Papers?
Maybe. We'll see. No promises here πŸ€“

# Contact Details
I'm not great at self-marketing (to say the least) and don't have any social media accounts. If you'd like to reach out to me, you can email me at [email protected]. Please note that this email might receive more messages than I can handle, so I apologize in advance if I can't respond to everyone.

# Versions and QUANTS
- Base model: [FP16](https://huggingface.co/SicariusSicariiStuff/Zion_Alpha)
- Instruction tuned: [FP16](https://huggingface.co/SicariusSicariiStuff/Zion_Alpha_Instruction_Tuned) | [GGUF](https://huggingface.co/SicariusSicariiStuff/Zion_Alpha_Instruction_Tuned_GGUF)


# Model architecture
Based on Mistral 7B. I didn't even bother to alter the tokenizer.

# The recommended prompt setting is Debug-deterministic:
```
temperature: 1
top_p: 1
top_k: 1
typical_p: 1
min_p: 1
repetition_penalty: 1
```

# The recommended instruction template is Mistral:
```
{%- for message in messages %}
    {%- if message['role'] == 'system' -%}
        {{- message['content'] -}}
    {%- else -%}
        {%- if message['role'] == 'user' -%}
            {{-'[INST] ' + message['content'].rstrip() + ' [/INST]'-}}
        {%- else -%}
            {{-'' + message['content'] + '</s>' -}}
        {%- endif -%}
    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{-''-}}
{%- endif -%}
```
# English to hebrew example:


<div align="center">
  <b style="font-size: 40px;">Zion_Alpha English to Hebrew example</b>


</div>


<img src="https://i.imgur.com/JnTuawF.png" alt="Zion_Alpha" style="width: 40%; min-width: 600px; display: block; margin: auto;">


# English to hebrew example:


<div align="center">
  <b style="font-size: 40px;">Zion_Alpha Hebrew to English example</b>


</div>


<img src="https://i.imgur.com/Wm2igLJ.png" alt="Zion_Alpha" style="width: 40%; min-width: 600px; display: block; margin: auto;">


<div align="center">
  <b style="font-size: 30px;">Unscripted video: live zero shot demonstration at story writing capabilities in Hebrew</b>

[![Zion_Alpha Story writing](https://img.youtube.com/vi/YYKeovnS0do/0.jpg)](https://www.youtube.com/watch?v=YYKeovnS0do)
</div>

<div align="center">
  <b style="font-size: 30px;">Zion_Alpha VS Mistral 'Hebrew' Live & unscripted in real time</b>

[![Zion_Alpha Story writing](https://img.youtube.com/vi/YYKeovnS0do/0.jpg)](https://www.youtube.com/watch?v=DQFtx8M2txc)
</div>

<div align="center">
  <b style="font-size: 30px;">Zion_Alpha VS Mistral 'Hebrew' Live & unscripted in real time Long text translation</b>

[![Zion_Alpha Story writing](https://img.youtube.com/vi/YYKeovnS0do/0.jpg)](https://www.youtube.com/watch?v=w5fz3Ot6tH8)
</div>

### History
The model was originally trained about 2 month after Mistral (v0.1) was released.
As of 04 June 2024, Zion_Alpha got the **Highest SNLI score in the world** among open source models in Hebrew, surpassing most of the models by a huge margin. (**84.05** score)
<img src="https://i.imgur.com/7HokS5w.png" alt="Zion_Alpha SNLI Score" style="width: 80%; min-width: 700px; display: block; margin: auto;">

### Support
<img src="https://i.imgur.com/0lHHN95.png" alt="GPUs too expensive" style="width: 10%; min-width: 100px; display: block; margin: left;">

- [My Ko-fi page](https://ko-fi.com/sicarius) ALL donations will go for research resources and compute, every bit counts πŸ™πŸ»
- [My Patreon](https://patreon.com/TenebraAI) ALL donations will go for research resources and compute, every bit counts πŸ™πŸ»