bjoernp commited on
Commit
f7e3d6f
1 Parent(s): e9f8c97

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +207 -0
README.md ADDED
@@ -0,0 +1,207 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - LeoLM/OpenSchnabeltier
4
+ - OpenAssistant/OASST-DE
5
+ - FreedomIntelligence/alpaca-gpt4-deutsch
6
+ - FreedomIntelligence/evol-instruct-deutsch
7
+ - LeoLM/German_Poems
8
+ - LeoLM/German_Songs
9
+ language:
10
+ - en
11
+ - de
12
+ library_name: transformers
13
+ pipeline_tag: text-generation
14
+ ---
15
+ # LAION LeoLM: **L**inguistically **E**nhanced **O**pen **L**anguage **M**odel
16
+ Meet LeoLM, the first open and commercially available German Foundation Language Model built on Llama-2.
17
+ Our models extend Llama-2's capabilities into German through continued pretraining on a large corpus of German-language and mostly locality specific text.
18
+ Thanks to a compute grant at HessianAI's new supercomputer **42**, we release three foundation models trained with 8k context length,
19
+ [`LeoLM/leo-mistral-hessianai-7b`](https://huggingface.co/LeoLM/leo-mistral-hessianai-7b), [`LeoLM/leo-hessianai-7b`](https://huggingface.co/LeoLM/leo-hessianai-7b) and [`LeoLM/leo-hessianai-13b`](https://huggingface.co/LeoLM/leo-hessianai-13b) under the [Llama-2 community license](https://huggingface.co/meta-llama/Llama-2-70b/raw/main/LICENSE.txt) (70b also coming soon! 👀).
20
+ With this release, we hope to bring a new wave of opportunities to German open-source and commercial LLM research and accelerate adoption.
21
+ Read our [blog post](https://laion.ai/blog/leo-lm/) or our paper (preprint coming soon) for more details!
22
+
23
+ *A project by Björn Plüster and Christoph Schuhmann in collaboration with LAION and HessianAI.*
24
+
25
+ ## LeoLM Chat
26
+ `LeoLM/leo-mistral-hessianai-7b-chat` is a German chat model built on our foundation model `LeoLM/leo-mistral-hessianai-7b` and finetuned on a selection of German instruction datasets.
27
+ The model performs exceptionally well on writing, explanation and discussion tasks but struggles somewhat with math and advanced reasoning. See our MT-Bench-DE scores:
28
+ ```
29
+ {
30
+ "first_turn": 6.1,
31
+ "second_turn": 4.7,
32
+ "categories": {
33
+ "writing": 6.8,
34
+ "roleplay": 6.35,
35
+ "reasoning": 3.3,
36
+ "math": 2.75,
37
+ "coding": 4.4,
38
+ "extraction": 4.5,
39
+ "stem": 6.85,
40
+ "humanities": 8.25
41
+ },
42
+ "average": 5.4
43
+ }
44
+ ```
45
+
46
+ ## Model Details
47
+
48
+ - **Finetuned from:** [LeoLM/leo-mistral-hessianai-7b](https://huggingface.co/LeoLM/leo-hessianai-7b)
49
+ - **Model type:** Causal decoder-only transformer language model
50
+ - **Language:** English and German
51
+ - **Demo:** [Web Demo coming soon !]()
52
+ - **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0.html)
53
+ - **Contact:** [LAION Discord](https://discord.com/invite/eq3cAMZtCC) or [Björn Plüster](mailto:bjoern.pl@outlook.de)
54
+
55
+
56
+ ## Use in 🤗Transformers
57
+ First install direct dependencies:
58
+ ```
59
+ pip install transformers torch sentencepiece
60
+ ```
61
+ If you want faster inference using flash-attention2, you need to install these dependencies:
62
+ ```bash
63
+ pip install packaging ninja
64
+ pip install flash-attn
65
+ ```
66
+ Then load the model in transformers:
67
+ ```python
68
+ from transformers import pipeline
69
+ import torch
70
+
71
+ system_prompt = """<|im_start|>system
72
+ Dies ist eine Unterhaltung zwischen einem intelligenten, hilfsbereitem KI-Assistenten und einem Nutzer.
73
+ Der Assistent gibt ausführliche, hilfreiche und ehrliche Antworten.<|im_end|>
74
+
75
+ """
76
+ prompt_format = "<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
77
+ prompt = "Erkläre mir wie die Fahrradwegesituation in Hamburg ist."
78
+
79
+ generator = pipeline(model="LeoLM/leo-mistral-hessianai-7b-chat", device="cuda", torch_dtype=torch.float16, use_flash_attention_2=True) # True for flash-attn2 else False
80
+ print(generator(prompt_format.format(prompt=prompt), do_sample=True, top_p=0.95, max_length=8192))
81
+ ```
82
+
83
+ "*Als KI kann ich keine persönlichen Beobachtungen teilen, aber ich kann einige allgemeine Informationen zur Fahrradwegesituation in Hamburg liefern. Im Vergleich zu vielen anderen großen Städten hat Hamburg eine hohe Anzahl von Fahrradfahrern und nimmt seine Verantwortung für nachhaltige Verkehrsmittel sehr ernst. Es gibt viele Fahrradwege und separate Fahrspuren, die Radfahrern ermöglichen, zusammen mit dem Autoverkehr zu fahren. Diese Fahrradspuren sind oft mit Markierungen gekennzeichnet und durch physische Trennungen von anderen Fahrspuren abgegrenzt. Darüber hinaus gibt es viele Fahrradstraßen, auf denen Radfahrer Vorfahrt haben und Autos langsamer fahren müssen.*
84
+
85
+ *In einigen städtischen Gebieten können Fahrradwege jedoch eng oder überfüllt sein, besonders während der Stoßzeiten. Es gibt auch viele Kreuzungen, an denen Radfahrer anhalten und auf Grün warten müssen, ähnlich wie Autofahrer. Insgesamt ist die Fahrradinfrastruktur in Hamburg ziemlich gut, aber wie überall gibt es immer Raum für Verbesserungen.*"
86
+
87
+ ## Prompting / Prompt Template
88
+
89
+ Prompt dialogue template (ChatML format):
90
+
91
+ ```
92
+ """
93
+ <|im_start|>system
94
+ {system_message}<|im_end|>
95
+ <|im_start|>user
96
+ {prompt}<|im_end|>
97
+ <|im_start|>assistant
98
+ """
99
+ ```
100
+
101
+ The model input can contain multiple conversation turns between user and assistant, e.g.
102
+ ```
103
+ <|im_start|>user
104
+ {prompt 1}<|im_end|>
105
+ <|im_start|>assistant
106
+ {reply 1}<|im_end|>
107
+ <|im_start|>user
108
+ {prompt 2}<|im_end|>
109
+ <|im_start|>assistant
110
+ (...)
111
+ ```
112
+
113
+ ## Ethical Considerations and Limitations
114
+
115
+ LeoLM has been tested in English and German, and has not covered, nor could it cover all scenarios.
116
+ For these reasons, as with all LLMs, the potential outputs of `LeoLM/leo-mistral-hessianai-7b-chat` cannot be predicted
117
+ in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses
118
+ to user prompts. Therefore, before deploying any applications of `LeoLM/leo-mistral-hessianai-7b-chat`, developers should
119
+ perform safety testing and tuning tailored to their specific applications of the model.
120
+
121
+ Please see Meta's [Responsible Use Guide](https://ai.meta.com/llama/responsible-use-guide/).
122
+
123
+ ## Finetuning Details
124
+
125
+ | Hyperparameter | Value |
126
+ |---|---|
127
+ | Num epochs | 4 |
128
+ | Examples per epoch | 131214 |
129
+ | Global batch size | 256 |
130
+ | Learning rate | 1e-5 |
131
+ | Warmup steps | 100 |
132
+ | LR scheduler | Cosine |
133
+ | Adam betas | (0.9, 0.95) |
134
+
135
+
136
+ ## Dataset Details
137
+ ```
138
+ ## Stats for 'Subset of OpenAssistant/OASST-DE' (3534 samples (100.0%))
139
+ -----------------
140
+ Accepted: 3534/3534 (100.0%)
141
+ Accepted tokens: 2259302
142
+ Skipped: 0 (0.0%)
143
+ Min tokens per sample: 29
144
+ Max tokens per sample: 2484
145
+ Avg tokens per sample: 639.3044708545557
146
+ -----------------
147
+
148
+ ## Stats for 'Subset of FreedomIntelligence/evol-instruct-deutsch' (57841 samples (100.0%))
149
+ -----------------
150
+ Accepted: 57841/57841 (100.0%)
151
+ Accepted tokens: 42958192
152
+ Skipped: 0 (0.0%)
153
+ Min tokens per sample: 33
154
+ Max tokens per sample: 5507
155
+ Avg tokens per sample: 742.6944900675991
156
+ -----------------
157
+
158
+ ## Stats for 'Subset of FreedomIntelligence/alpaca-gpt4-deutsch' (48969 samples (100.0%))
159
+ -----------------
160
+ Accepted: 48969/48969 (100.0%)
161
+ Accepted tokens: 13372005
162
+ Skipped: 0 (0.0%)
163
+ Min tokens per sample: 19
164
+ Max tokens per sample: 1359
165
+ Avg tokens per sample: 273.07082031489307
166
+ -----------------
167
+
168
+ ## Stats for 'Subset of LeoLM/OpenSchnabeltier' (21314 samples (100.0%))
169
+ -----------------
170
+ Accepted: 21314/21314 (100.0%)
171
+ Accepted tokens: 8134690
172
+ Skipped: 0 (0.0%)
173
+ Min tokens per sample: 25
174
+ Max tokens per sample: 1202
175
+ Avg tokens per sample: 381.65947264708643
176
+ -----------------
177
+
178
+ ## Stats for 'Subset of LeoLM/German_Poems' (490 samples (100.0%))
179
+ -----------------
180
+ Accepted: 490/490 (100.0%)
181
+ Accepted tokens: 618642
182
+ Skipped: 0 (0.0%)
183
+ Min tokens per sample: 747
184
+ Max tokens per sample: 1678
185
+ Avg tokens per sample: 1262.534693877551
186
+ -----------------
187
+
188
+ ## Stats for 'Subset of LeoLM/German_Songs' (392 samples (100.0%))
189
+ -----------------
190
+ Accepted: 392/392 (100.0%)
191
+ Accepted tokens: 187897
192
+ Skipped: 0 (0.0%)
193
+ Min tokens per sample: 231
194
+ Max tokens per sample: 826
195
+ Avg tokens per sample: 479.3290816326531
196
+ -----------------
197
+
198
+ ## Stats for 'total' (132540 samples (100.0%))
199
+ -----------------
200
+ Accepted: 132540/132540 (100.0%)
201
+ Accepted tokens: 67530728
202
+ Skipped: 0 (0.0%)
203
+ Min tokens per sample: 19
204
+ Max tokens per sample: 5507
205
+ Avg tokens per sample: 509.51205673758864
206
+ -----------------
207
+ ```