johannhartmann commited on
Commit
5e5ab2f
1 Parent(s): b009c51

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +137 -0
README.md ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ - it
6
+ pipeline_tag: text-generation
7
+ ---
8
+
9
+ ![image/png](https://huggingface.co/datasets/malteos/images/resolve/main/occiglot.medium.png)
10
+
11
+ # Occiglot-7B-it-en-Instruct
12
+
13
+ > A [polyglot](https://en.wikipedia.org/wiki/Multilingualism#In_individuals) language model for the [Occident](https://en.wikipedia.org/wiki/Occident).
14
+ >
15
+
16
+ **Occiglot-7B-EU5-Instruct** is a the instruct version of [occiglot-7b-it-en](https://huggingface.co/occiglot/occiglot-7b-it-en), a generative language model with 7B parameters supporting the top-5 EU languages (English, Spanish, French, German, and Italian) and trained by the [Occiglot Research Collective](https://occiglot.github.io/occiglot/).
17
+ It was trained on 160M tokens of additional multilingual and code instructions.
18
+ Note that the model was not safety aligned and might generate problematic outputs.
19
+
20
+ This is the first release of an ongoing open research project for multilingual language models.
21
+ If you want to train a model for your own language or are working on evaluations, please contact us or join our [Discord server](https://discord.gg/wUpvYs4XvM). **We are open for collaborations!**
22
+
23
+
24
+ ### Model details
25
+
26
+ - **Instruction tuned from:** [occiglot-7b-it-en](https://huggingface.co/occiglot/occiglot-7b-it-en)
27
+ - **Model type:** Causal decoder-only transformer language model
28
+ - **Languages:** English, Italian, and code.
29
+ - **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0.html)
30
+ - **Compute resources:** [DFKI cluster](https://www.dfki.de/en/web)
31
+ - **Contributors:** Manuel Brack, Patrick Schramowski, Pedro Ortiz, Malte Ostendorff, Fabio Barth, Georg Rehm, Kristian Kersting
32
+ - **Research labs:** [Occiglot](https://occiglot.github.io/occiglot/) with support from [SAINT](https://www.dfki.de/en/web/research/research-departments/foundations-of-systems-ai) and [SLT](https://www.dfki.de/en/web/research/research-departments/speech-and-language-technology)
33
+ - **Contact:** [Discord](https://discord.gg/wUpvYs4XvM)
34
+
35
+ ### How to use
36
+
37
+ The model was trained using the chatml instruction template. You can use the transformers chat template feature for interaction.
38
+ Since the generation relies on some randomness, we
39
+ set a seed for reproducibility:
40
+
41
+ ```python
42
+ >>> from transformers import AutoTokenizer, MistralForCausalLM, set_seed
43
+ >>> tokenizer = AutoTokenizer.from_pretrained("occiglot/occiglot-7b-eu5-instruct")
44
+ >>> model = MistralForCausalLM.from_pretrained('occiglot/occiglot-7b-eu5-instruct') # You may want to use bfloat16 and/or move to GPU here
45
+ >>> set_seed(42)
46
+ >>> messages = [
47
+ >>> {"role": "system", 'content': 'You are a helpful assistant. Please give short and concise answers.'},
48
+ >>> {"role": "user", "content": "chi è il primo ministro italiano?"},
49
+ >>> ]
50
+ >>> tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_dict=False, return_tensors='pt',)
51
+ >>> set_seed(42)
52
+ >>> outputs = model.generate(tokenized_chat.to('cuda'), max_new_tokens=200,)
53
+ >>> tokenizer.decode(out[0][len(tokenized_chat[0]):])
54
+ 'Il primo ministro italiano è attualmente Giorgia Meloni, presidente di Fratelli d'Italia, un partito politico di estrema destra.'
55
+ ```
56
+
57
+ ## Dataset
58
+
59
+ The training data was split evenly amongst the 5 languages based on the total number of tokens. We would like to thank Disco Research and Björn Plüster for making their dataset available to us.
60
+
61
+ **English and Code**
62
+ - [Open-Hermes-2B](https://huggingface.co/datasets/teknium/OpenHermes-2.5)
63
+
64
+ **Italian**
65
+ - [Quora-IT-Baize](https://huggingface.co/datasets/andreabac3/Quora-Italian-Fauno-Baize)
66
+ - [Stackoverflow-IT-Vaize](https://huggingface.co/datasets/andreabac3/StackOverflow-Italian-Fauno-Baize)
67
+ - [Camoscio](https://huggingface.co/datasets/teelinsan/camoscio_cleaned)
68
+ - [OASST-2](https://huggingface.co/datasets/OpenAssistant/oasst2) (Italian subset)
69
+ - [Aya-Dataset](https://huggingface.co/datasets/CohereForAI/aya_dataset) (Italian subset)
70
+
71
+ ## Training settings
72
+
73
+ - Full instruction fine-tuning on 8xH100.
74
+ - 0.6 - 4 training epochs (depending on dataset sampling).
75
+ - Framework: [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl)
76
+ - Precision: bf16
77
+ - Optimizer: AdamW
78
+ - Global batch size: 128 (with 8192 context length)
79
+ - Cosine Annealing with Warmup
80
+
81
+
82
+ ## Tokenizer
83
+
84
+ Tokenizer is unchanged from [Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1).
85
+
86
+ ## Evaluation
87
+
88
+ Preliminary evaluation results can be found below.
89
+ Please note that the non-English results are based on partially machine-translated datasets and English prompts ([Belebele](https://huggingface.co/datasets/facebook/belebele) and [Okapi framework](https://github.com/nlp-uoregon/Okapi)) and thus should be interpreted with caution, e.g., biased towards English model performance.
90
+ Currently, we are working on more suitable benchmarks for Spanish, French, German, and Italian.
91
+
92
+ <details>
93
+ <summary>Evaluation results</summary>
94
+
95
+ ### English
96
+
97
+ | | arc_challenge | belebele | hellaswag | mmlu | truthfulqa | avg |
98
+ |:-------------------------------------|----------------:|-----------:|------------:|---------:|-------------:|---------:|
99
+ | occiglot/occiglot-7b-eu5 | 0.530717 | 0.726667 | 0.789882 | 0.531904 | 0.403678 | 0.59657 |
100
+ | occiglot/occiglot-7b-eu5-instruct | 0.558874 | 0.746667 | 0.799841 | 0.535109 | 0.449034 | 0.617905 |
101
+ | occiglot/occiglot-7b-it-en | 0.580205 | 0.774444 | 0.804222 | 0.578977 | 0.412786 | 0.630127 |
102
+ | occiglot/occiglot-7b-it-en-instruct | 0.609215 | 0.82 | 0.809301 | 0.578835 | 0.479562 | 0.659383 |
103
+ | galatolo/cerbero-7b | 0.613481 | 0.827778 | 0.810396 | 0.600484 | 0.480911 | 0.66661 |
104
+ | mistralai/Mistral-7B-v0.1 | 0.612628 | 0.844444 | 0.834097 | 0.624555 | 0.426201 | 0.668385 |
105
+ | mistralai/Mistral-7B-Instruct-v0.2 | 0.637372 | 0.824444 | 0.846345 | 0.59201 | 0.668116 | 0.713657 |
106
+
107
+
108
+ ### Italian
109
+
110
+ | | arc_challenge_it | belebele_it | hellaswag_it | mmlu_it | truthfulqa_it | avg |
111
+ |:-------------------------------------|-------------------:|--------------:|---------------:|----------:|----------------:|---------:|
112
+ | occiglot/occiglot-7b-eu5 | 0.501283 | 0.652222 | 0.700533 | 0 | 0.252874 | 0.421382 |
113
+ | occiglot/occiglot-7b-eu5-instruct | 0.516681 | 0.661111 | 0.71326 | 0 | 0.295019 | 0.437214 |
114
+ | occiglot/occiglot-7b-it-en | 0.536356 | 0.684444 | 0.694768 | 0 | 0.247765 | 0.432667 |
115
+ | occiglot/occiglot-7b-it-en-instruct | 0.545766 | 0.717778 | 0.713804 | 0 | 0.303959 | 0.456261 |
116
+ | galatolo/cerbero-7b | 0.522669 | 0.717778 | 0.631567 | 0 | 0.302682 | 0.434939 |
117
+ | mistralai/Mistral-7B-v0.1 | 0.502139 | 0.734444 | 0.630371 | 0 | 0.264368 | 0.426264 |
118
+ | mistralai/Mistral-7B-Instruct-v0.2 | 0.519247 | 0.703333 | 0.6394 | 0 | 0.349936 | 0.442383 |
119
+
120
+
121
+ </details>
122
+
123
+ ## Acknowledgements
124
+
125
+ The pre-trained model training was supported by a compute grant at the [42 supercomputer](https://hessian.ai/) which is a central component in the development of [hessian AI](https://hessian.ai/), the [AI Innovation Lab](https://hessian.ai/infrastructure/ai-innovationlab/) (funded by the [Hessian Ministry of Higher Education, Research and the Art (HMWK)](https://wissenschaft.hessen.de) & the [Hessian Ministry of the Interior, for Security and Homeland Security (HMinD)](https://innen.hessen.de)) and the [AI Service Centers](https://hessian.ai/infrastructure/ai-service-centre/) (funded by the [German Federal Ministry for Economic Affairs and Climate Action (BMWK)](https://www.bmwk.de/Navigation/EN/Home/home.html)).
126
+ The curation of the training data is partially funded by the [German Federal Ministry for Economic Affairs and Climate Action (BMWK)](https://www.bmwk.de/Navigation/EN/Home/home.html)
127
+ through the project [OpenGPT-X](https://opengpt-x.de/en/) (project no. 68GX21007D).
128
+
129
+
130
+ ## License
131
+
132
+ [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0.html)
133
+
134
+ ## See also
135
+
136
+ - https://huggingface.co/collections/occiglot/occiglot-eu5-7b-v01-65dbed502a6348b052695e01
137
+ - https://huggingface.co/NikolayKozloff/occiglot-7b-it-en-GGUF