Occiglot-7B-EU5-Instruct

A polyglot language model for the Occident.

Occiglot-7B-EU5-Instruct is a the instruct version of occiglot-7b-eu5, 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. It was trained on 400M tokens of additional multilingual and code instructions. Note that the model was not safety aligned and might generate problematic outputs.

This is the first release of an ongoing open research project for multilingual language models. If you want to train a model for your own language or are working on evaluations, please contact us or join our Discord server. We are open for collaborations!

Model details

Instruction tuned from: occiglot-7b-eu5
Model type: Causal decoder-only transformer language model
Languages: English, Spanish, French, German, Italian, and code.
License: Apache 2.0
Compute resources: DFKI cluster
Contributors: Manuel Brack, Patrick Schramowski, Pedro Ortiz, Malte Ostendorff, Fabio Barth, Georg Rehm, Kristian Kersting
Research labs: Occiglot with support from SAINT and SLT
Contact: Discord

How to use

The model was trained using the chatml instruction template. You can use the transformers chat template feature for interaction. Since the generation relies on some randomness, we set a seed for reproducibility:

>>> from transformers import AutoTokenizer, MistralForCausalLM, set_seed
>>> tokenizer = AutoTokenizer.from_pretrained("occiglot/occiglot-7b-eu5-instruct")
>>> model = MistralForCausalLM.from_pretrained('occiglot/occiglot-7b-eu5-instruct')  # You may want to use bfloat16 and/or move to GPU here
>>> set_seed(42)
>>> messages = [
>>>    {"role": "system", 'content': 'You are a helpful assistant. Please give short and concise answers.'},
>>>    {"role": "user", "content": "Wer ist der deutsche Bundeskanzler?"},
>>> ]
>>> tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_dict=False, return_tensors='pt',)
>>> set_seed(42)
>>> outputs = model.generate(tokenized_chat.to('cuda'), max_new_tokens=200,)
>>> tokenizer.decode(out[0][len(tokenized_chat[0]):])
'Der deutsche Bundeskanzler ist Olaf Scholz.'

Dataset

The training data was split evenly amongst the 5 languages based on the total number of tokens. We would like to thank Disco Research, Jan Philipp Harries, and Björn Plüster for making their dataset available to us.

English and Code

Open-Hermes-2B

German

DiscoLM German Dataset includes the publicly available germanrag dataset
OASST-2 (German subset)
Aya-Dataset (German subset)

Spanish

Mentor-ES
Squad-es
OASST-2 (Spanish subset)
Aya-Dataset (Spanish subset)

French

Bactrian-X (French subset)
AI-Society Translated (French subset)
GT-Dorimiti
OASST-2 (French subset)
Aya-Dataset (French subset)

Italian

Quora-IT-Baize
Stackoverflow-IT-Vaize
Camoscio
OASST-2 (Italian subset)
Aya-Dataset (Italian subset)

Training settings

Full instruction fine-tuning on 8xH100.
0.6 - 4 training epochs (depending on dataset sampling).
Framework: axolotl
Precision: bf16
Optimizer: AdamW
Global batch size: 128 (with 8192 context length)
Cosine Annealing with Warmup

Tokenizer

Tokenizer is unchanged from Mistral-7B-v0.1.

Evaluation

Preliminary evaluation results can be found below. Please note that the non-English results are based on partially machine-translated datasets and English prompts (Belebele and Okapi framework) and thus should be interpreted with caution, e.g., biased towards English model performance. Currently, we are working on more suitable benchmarks for Spanish, French, German, and Italian.

Evaluation results

All 5 Languages

	avg	arc_challenge	belebele	hellaswag	mmlu	truthfulqa
Occiglot-7b-eu5	0.516895	0.508109	0.675556	0.718963	0.402064	0.279782
Occiglot-7b-eu5-instruct	0.537799	0.53632	0.691111	0.731918	0.405198	0.32445
Occiglot-7b-de-en	0.518337	0.496297	0.715111	0.669034	0.412545	0.298697
Occiglot-7b-de-en-instruct	0.543173	0.530826	0.745778	0.67676	0.411326	0.351176
Occiglot-7b-it-en	0.513221	0.500564	0.694444	0.668099	0.413528	0.289469
Occiglot-7b-it-en-instruct	0.53721	0.523128	0.726667	0.683414	0.414918	0.337927
Occiglot-7b-fr-en	0.509209	0.496806	0.691333	0.667475	0.409129	0.281303
Occiglot-7b-fr-en-instruct	0.52884	0.515613	0.723333	0.67371	0.413024	0.318521
Occiglot-7b-es-en	0.483388	0.482949	0.606889	0.653902	0.398922	0.274277
Occiglot-7b-es-en-instruct	0.504023	0.494576	0.65	0.670847	0.406176	0.298513
Leo-mistral-hessianai-7b	0.484806	0.462103	0.653556	0.642242	0.379208	0.28692
Claire-mistral-7b-0.1	0.514226	0.502773	0.705111	0.666871	0.412128	0.284245
Lince-mistral-7b-it-es	0.543427	0.540222	0.745111	0.692931	0.426241	0.312629
Cerbero-7b	0.532385	0.513714	0.743111	0.654061	0.427566	0.323475
Mistral-7b-v0.1	0.547111	0.528937	0.768444	0.682516	0.448253	0.307403
Mistral-7b-instruct-v0.2	0.56713	0.547228	0.741111	0.69455	0.422501	0.430262

English

	avg	arc_challenge	belebele	hellaswag	mmlu	truthfulqa
Occiglot-7b-eu5	0.59657	0.530717	0.726667	0.789882	0.531904	0.403678
Occiglot-7b-eu5-instruct	0.617905	0.558874	0.746667	0.799841	0.535109	0.449
Leo-mistral-hessianai-7b	0.600949	0.522184	0.736667	0.777833	0.538812	0.429248
Mistral-7b-v0.1	0.668385	0.612628	0.844444	0.834097	0.624555	0.426201
Mistral-7b-instruct-v0.2	0.713657	0.637372	0.824444	0.846345	0.59201	0.668116

German

	avg	arc_challenge_de	belebele_de	hellaswag_de	mmlu_de	truthfulqa_de
Occiglot-7b-eu5	0.508311	0.493584	0.646667	0.666631	0.483406	0.251269
Occiglot-7b-eu5-instruct	0.531506	0.529512	0.667778	0.685205	0.488234	0.286802
Occiglot-7b-de-en	0.540085	0.50556	0.743333	0.67421	0.514633	0.26269
Occiglot-7b-de-en-instruct	0.566474	0.54491	0.772222	0.688407	0.515915	0.310914
Leo-mistral-hessianai-7b	0.517766	0.474765	0.691111	0.682109	0.488309	0.252538
Mistral-7b-v0.1	0.527957	0.476476	0.738889	0.610589	0.529567	0.284264
Mistral-7b-instruct-v0.2	0.535215	0.485885	0.688889	0.622438	0.501961	0.376904

Spanish

	avg	arc_challenge_es	belebele_es	hellaswag_es	mmlu_es	truthfulqa_es
Occiglot-7b-eu5	0.533194	0.508547	0.676667	0.725411	0.499325	0.25602
Occiglot-7b-eu5-instruct	0.548155	0.535043	0.68	0.737039	0.503525	0.285171
Occiglot-7b-es-en	0.527264	0.529915	0.627778	0.72253	0.512749	0.243346
Occiglot-7b-es-en-instruct	0.5396	0.545299	0.636667	0.734372	0.524374	0.257288
Lince-mistral-7b-it-es	0.547212	0.52906	0.721111	0.687967	0.512749	0.285171
Mistral-7b-v0.1	0.554817	0.528205	0.747778	0.672712	0.544023	0.281369
Mistral-7b-instruct-v0.2	0.568575	0.54188	0.73	0.685406	0.511699	0.373891

French

	avg	arc_challenge_fr	belebele_fr	hellaswag_fr	mmlu_fr	truthfulqa_fr
Occiglot-7b-eu5	0.525017	0.506416	0.675556	0.712358	0.495684	0.23507
Occiglot-7b-eu5-instruct	0.554216	0.541488	0.7	0.724245	0.499122	0.306226
Occiglot-7b-fr-en	0.542903	0.532934	0.706667	0.718891	0.51333	0.242694
Occiglot-7b-fr-en-instruct	0.567079	0.542344	0.752222	0.72553	0.52051	0.29479
Claire-mistral-7b-0.1	0.515127	0.486741	0.694444	0.642964	0.479566	0.271919
Cerbero-7b	0.526044	0.462789	0.735556	0.624438	0.516462	0.290978
Mistral-7b-v0.1	0.558129	0.525235	0.776667	0.66481	0.543121	0.280813
Mistral-7b-instruct-v0.2	0.575821	0.551754	0.758889	0.67916	0.506837	0.382465

Italian

	avg	arc_challenge_it	belebele_it	hellaswag_it	truthfulqa_it
Occiglot-7b-eu5	0.421382	0.501283	0.652222	0.700533	0.252874
Occiglot-7b-eu5-instruct	0.437214	0.516681	0.661111	0.71326	0.295019
Occiglot-7b-it-en	0.432667	0.536356	0.684444	0.694768	0.247765
Occiglot-7b-it-en-instruct	0.456261	0.545766	0.717778	0.713804	0.303959
Cerbero-7b	0.434939	0.522669	0.717778	0.631567	0.302682
Mistral-7b-v0.1	0.426264	0.502139	0.734444	0.630371	0.264368
Mistral-7b-instruct-v0.2	0.442383	0.519247	0.703333	0.6394	0.349936

Acknowledgements

The pre-trained model training was supported by a compute grant at the 42 supercomputer which is a central component in the development of hessian AI, the AI Innovation Lab (funded by the Hessian Ministry of Higher Education, Research and the Art (HMWK) & the Hessian Ministry of the Interior, for Security and Homeland Security (HMinD)) and the AI Service Centers (funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK)). The curation of the training data is partially funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK) through the project OpenGPT-X (project no. 68GX21007D).

License

Apache 2.0

mayflowergmbh
/

occiglot-7b-eu5-instruct-hf