Model Card for Teuken-7B-instruct-commercial-v0.4

Teuken-7B-instruct-commercial-v0.4 is an instruction-tuned 7B parameter multilingual large language model (LLM) pre-trained with 4T tokens in all official 24 European languages and released under Apache 2.0 in the research project OpenGPT-X. The base model Teuken-7B-base-v0.4 is available on request 📧 contact@opengpt-x.de.

Model Description

• Developed by: Fraunhofer, Forschungszentrum Jülich, TU Dresden, DFKI
• Funded by: German Federal Ministry of Economics and Climate Protection (BMWK) in the context of the OpenGPT-X project
• Model type: Transformer based decoder-only model
• Language(s) (NLP): bg, cs, da, de, el, en, es, et, fi, fr, ga, hr, hu, it, lt, lv, mt, nl, pl, pt, ro, sk, sl, sv
• Shared by: OpenGPT-X

Uses

Teuken-7B-instruct-commercial-v0.4 is intended for commercial and research use in all official 24 European languages. Since Teuken-7B-instruct-commercial-v0.4 focuses on covering all 24 EU languages, it renders more stable results across these languages and better reflects European values in its answers than English-centric models. It is therefore specialized for use in multilingual tasks.

Disclaimer Toxic Content:

This Large Language Model (LLM) may generate content that is inappropriate, offensive, or harmful. While the dataset has been heavily filtered to minimize such outputs, the model may still produce text that is biased or toxic due to the large scale and diverse nature of the data.

Out-of-Scope Use

The model is not intended for use in math and coding tasks.

Bias, Risks, and Limitations

Teuken-7B-instruct-commercial-v0.4 is an instruction-tuned version of Teuken-7B-base-v0.4 (which is available on request 📧 contact@opengpt-x.de) that is not completely free from biases and hallucinations.

How to Get Started with the Model

Usage

The model requires transformers, sentencepiece, and the torch library. After installation, here's an example of how to use the model:

As this model is a fine-tuned model, it must be used with the provided prompt template. Using the model without the prompt template is not intended and is not recommended. The prompt template is defined as follows:

user="Hi!"
lang_code = "DE"
system_messages={
            "EN": "A chat between a human and an artificial intelligence assistant."
            " The assistant gives helpful and polite answers to the human's questions.",
            "DE": "Ein Gespräch zwischen einem Menschen und einem Assistenten mit künstlicher Intelligenz."
            " Der Assistent gibt hilfreiche und höfliche Antworten auf die Fragen des Menschen.",
        }
 
prompt = f"System: {system_messages[lang_code]}\nUser: {user}\nAssistant:"

The prompt template is also directly integrated in the Tokenizer and can be used as follows:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_name = "openGPT-X/Teuken-7B-instruct-commercial-v0.4"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
model = model.to(device).eval()
tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    use_fast=False,
    trust_remote_code=True,
)
messages = [{"role": "User", "content": "Wer bist du?"}]
prompt_ids = tokenizer.apply_chat_template(messages, chat_template="DE", tokenize=True, add_generation_prompt=True, return_tensors="pt")
prediction = model.generate(
    prompt_ids.to(model.device),
    max_length=512,
    do_sample=True,
    top_k=50,
    top_p=0.95,
    temperature=0.7,
    num_return_sequences=1,
)
prediction_text = tokenizer.decode(prediction[0].tolist())
print(prediction_text)

This example demonstrates how to load the model and tokenizer, prepare input, generate text, and print the result.

Training Details

Pre-Training Data

Teuken-7B-base-v0.4 was pre-trained on 4 trillion tokens of data from publicly available sources. The pretraining data has a cutoff of September 2023. More information is available in our preprint "Data Processing for the OpenGPT-X Model Family".

Instruction-Tuning Data

The model was fine-tuned on a collection of English- and German-focused instruction-tuning datasets which also contains instructions for 22 official European languages The dataset composition contains three types of data: multilingual data, English data, and translated German data

English data

• We only included a subsample of the OpenOrca dataset.
• To select instruction-tuning examples based on their quality, We calculated the reward scores of all English examples utilizing Starling-RM-7B-alpha (Apache-2.0 license)

We aim to include roughly the same amount of English examples as we have multilingual examples:

1. Add all multi-turn examples
2. Add entire code_alpaca dataset subset
3. For the remaining dataset subsets (open_orca, evol_instruct_143k, evol_instruct_70k, sharegpt_v3, ultrachat_200k), we add the samples with the highest reward scores so that each dataset subset contributes an equal amount of high-quality examples

German Data

As we aim for a German- and English-centric, European language dataset and due to the sparsity of large-scale German instruction-tuning data, we translated the English portion of the above-described dataset composition. For this, we applied the Alma-13B (MIT license) model. As code can be a problematic case for translation, we implemented a regex-based code detection functionality. With it, we exclude code snippets from translation and insert the code snippets after translation again. As the alpaca_code contains many code snippets not detectable by our regex-based code detection implementation, we included this part of the dataset from the translation.

Multilingual data

For multilingual data we include the 14 offical European languages contained in the aya_dataset and the 21 offical European languages contained in the translated_flan_cot dataset of the aya_collection.

Datasets and Licenses

Name	Language	License
Open-Orca/OpenOrca	EN	MIT
sahil2801/CodeAlpaca-20k	EN	CC-BY-4.0
WizardLM/WizardLM_evol_instruct_V2_196k	EN	MIT
WizardLM/WizardLM_evol_instruct_70k	EN	MIT
anon8231489123/ShareGPT_Vicuna_unfiltered	EN	Apache-2.0
HuggingFaceH4/ultrachat_200k	EN	MIT
CohereForAI/aya_dataset	Multilingual	Apache-2.0
CohereForAI/aya_collection	Multilingual	Apache-2.0
FreedomIntelligence/sharegpt-deutsch	DE	Apache-2.0
bjoernp/ultrachat_de	DE	MIT

Dataset contribution per language:

	total	de_freedomintelligence_sharegpt	de_ultrachat_de	translated_flan_cot	aya_dataset	ultrachat_200k_translated_to_de	sharegpt_v3_unfiltered_translated_to_de	evol_instruct_143k_translated_to_de	evol_instruct_70k_translated_to_de	open_orca_translated_to_de	ultrachat_200k	sharegpt_v3_unfiltered	code_alpaca	open_orca	evol_instruct_143k	evol_instruct_70k
BG	1909	0	0	1909	0	0	0	0	0	0	0	0	0	0	0	0
CS	1885	0	0	1885	0	0	0	0	0	0	0	0	0	0	0	0
DA	2001	0	0	1906	95	0	0	0	0	0	0	0	0	0	0	0
DE	77628	5818	898	1896	231	6940	37555	8116	8065	8109	0	0	0	0	0	0
ET	1901	0	0	1901	0	0	0	0	0	0	0	0	0	0	0	0
EL	2472	0	0	1881	591	0	0	0	0	0	0	0	0	0	0	0
ES	3800	0	0	1898	1902	0	0	0	0	0	0	0	0	0	0	0
EN	80806	0	0	0	0	0	0	0	0	0	6915	37600	12013	8074	8099	8105
FI	2598	0	0	1890	708	0	0	0	0	0	0	0	0	0	0	0
FR	3250	0	0	1890	1360	0	0	0	0	0	0	0	0	0	0	0
HU	1985	0	0	1892	93	0	0	0	0	0	0	0	0	0	0	0
MT	1918	0	0	1918	0	0	0	0	0	0	0	0	0	0	0	0
IT	2613	0	0	1910	703	0	0	0	0	0	0	0	0	0	0	0
LT	2800	0	0	1920	880	0	0	0	0	0	0	0	0	0	0	0
NL	3549	0	0	1905	1644	0	0	0	0	0	0	0	0	0	0	0
PL	3322	0	0	1909	1413	0	0	0	0	0	0	0	0	0	0	0
PT	3806	0	0	1897	1909	0	0	0	0	0	0	0	0	0	0	0
RO	1888	0	0	1888	0	0	0	0	0	0	0	0	0	0	0	0
GA	3069	0	0	1880	1189	0	0	0	0	0	0	0	0	0	0	0
SK	1922	0	0	1922	0	0	0	0	0	0	0	0	0	0	0	0
SL	1894	0	0	1894	0	0	0	0	0	0	0	0	0	0	0	0
SV	3160	0	0	1916	1244	0	0	0	0	0	0	0	0	0	0	0

Total across languages 210,176

Training Procedure

Instruction fined tuned version of Teuken-7B-base-v0.4. More information regarding the pre-training are available in our model preprint "Teuken-7B-Base & Teuken-7B-Instruct: Towards European LLMs".

Training Hyperparameters

• Training regime: bf16 mixed precision

Evaluation

Results on multilingual benchmarks for 21 European languages with instruction-tuned models

Model	Avg.	EU21-ARC	EU21-HeSw	EU21-TQA	EU21-MMLU
Meta-Llama-3.1-8B-Instruct	.563	.563	.579	.532	.576
Mistral-7B-Instruct-v0.3	.527	.530	.538	.548	.491
Salamandra-7B-Instruct	.543	.595	.637	.482	.459
Aya-23-8B	.485	.475	.535	.476	.455
Occiglot-7B-eu5-Instruct	.475	.484	.519	.471	.428
Pharia-1-LLM-7B-C-A	.417	.396	.438	.469	.366
Bloomz-7B1	.358	.316	.354	.461	.302
Teuken-7B-instruct-commercial-v0.4	.531	.569	.620	.503	.430

More information regarding the quality of our translated benchmarks are available in our Evaluation preprint "Towards Multilingual LLM Evaluation for European Languages". More evaluation results regarding Teuken-7B-instruct-research-v0.4 are available in our model preprint "Teuken-7B-Base & Teuken-7B-Instruct: Towards European LLMs".

The model was evaluated in 21 languages on ARC, GSM8K, HellaSwag, TruthfulQA, Translation and MMLU. Results can also be seen in the European LLM Leaderboard.

Technical Specifications

Model Architecture and Objective

Hyper-Parameter	Value
Training Objective	CLM
Activation Function	SwiGLU
Seq Length	4096
Position Embeddings	Rotary
Num Layers	32
Hidden Size	4096
FFN Hidden Size	13440
Num Attention Heads	32
Head Dim	128
Group Query Attention	yes
Num Query Groups	2
Normalization	RMSNorm
Learning rate	3e-4
Min learning rate	3e-5
Disable bias in linear	yes
Hidden dropout	0.0
Attention dropout	0.0
Optimizer	AdamW
Beta1	0.9
Beta2	0.95
Data-type	bf16
Recompute-activations	yes
Distributed-optimizers	yes

Compute Infrastructure

We trained our models on JUWELS Booster which consists of 936 compute nodes, each equipped with 4 NVIDIA A100 GPUs. The GPUs are hosted by AMD EPYC Rome CPUs. The compute nodes are connected with HDR-200 InfiniBand in a DragonFly+ topology.

Hardware

The configuration of JUWELS Booster compute nodes is the following:

CPU: AMD EPYC 7402 processor; 2 sockets, 24 cores per socket, SMT-2 (total: 2×24×2 = 96 threads) in NPS-4 1 configuration
Memory: 512 GB DDR4-3200 RAM (of which at least 20 GB is taken by the system software stack, including the file system); 256 GB per socket; 8 memory channels per socket (2 channels per NUMA domain)
GPU: 4 × NVIDIA A100 Tensor Core GPU with 40 GB; connected via NVLink3 to each other
Network: 4 × Mellanox HDR200 InfiniBand ConnectX 6 (200 Gbit/s each), HCA
Periphery: CPU, GPU, and network adapter are connected via 2 PCIe Gen 4 switches with 16 PCIe lanes going to each device (CPU socket: 2×16 lanes). PCIe switches are configured in synthetic mode.

Software

Megatron-LM

BibTeX:

If you find our model useful in your research, please consider citing our preprint:

@misc{ali2024teuken7bbaseteuken7binstructeuropean,
      title={Teuken-7B-Base & Teuken-7B-Instruct: Towards European LLMs}, 
      author={Mehdi Ali and Michael Fromm and Klaudia Thellmann and Jan Ebert and Alexander Arno Weber and Richard Rutmann and Charvi Jain and Max Lübbering and Daniel Steinigen and Johannes Leveling and Katrin Klug and Jasper Schulze Buschhoff and Lena Jurkschat and Hammam Abdelwahab and Benny Jörg Stein and Karl-Heinz Sylla and Pavel Denisov and Nicolo' Brandizzi and Qasid Saleem and Anirban Bhowmick and Lennard Helmer and Chelsea John and Pedro Ortiz Suarez and Malte Ostendorff and Alex Jude and Lalith Manjunath and Samuel Weinbach and Carolin Penke and Oleg Filatov and Shima Asaadi and Fabio Barth and Rafet Sifa and Fabian Küch and Andreas Herten and René Jäkel and Georg Rehm and Stefan Kesselheim and Joachim Köhler and Nicolas Flores-Herr},
      year={2024},
      eprint={2410.03730},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.03730}, 
}

Team

Data Team

Anirban Bhowmick (IAIS), Nicolo Brandizzi (IAIS), Lennard Helmer (IAIS), Benny Jörg Stein (IAIS), Karl-Heinz Sylla (IAIS), Pavel Denisov (IAIS), Qasid Saleem (IAIS), Johannes Leveling (IAIS), Hammam Abdelwahab (IAIS), Luzian Hahn (IIS), Farzad Naderi (IIS), Md Saiful Islam (IIS), Alexander Schwirjow (IIS), Pedro Ortiz Suarez (ex. DFKI), Malte Ostendorff (ex. DFKI)

Model-Training Team

Core contributors

Mehdi Ali (IAIS), Michael Fromm (IAIS), Jan Ebert (FZJ), Chelsea John (FZJ), Lena Jurkschat (TUD), Alexander Weber (IAIS)

Contributors:

Richard Rutmann (IAIS), Daniel Steinigen (IAIS), Lalith Manjunath (TUD), Carolin Penke (FZJ)

Evaluation Team

Core contributors

Klaudia Thellmann (TUD), Alex Jude (IAIS), Jasper Buschhoff (IAIS)

Contributors:

Shima Assadi (IIS), Fabio Barth (DFKI)

Management

Joachim Köhler (IAIS), Nicolas Flores-Herr (IAIS), Stefan Kesselheim (FZJ), Andreas Herten (FZJ), Georg Rehm (DFKI), René Jäkel (TUD), Fabian Küch (IIS), Nicole Hildebrandt (IAIS), Ines Wendler (IAIS)

We believe that collaboration is key to overcome the aforementioned limitations and thereby strengthening the European GenAI landscape. Because of this, the team invites researchers, developers, and AI enthusiasts to join and engage through various platforms. A Discord server has been created for community collaboration, offering a space for discussions on technical details, ideas, and direct interaction with developers. Additionally, resources like research publications and a European LLM Leaderboard provide insights into Teuken-7B’s performance and technical aspects. The OpenGPT-X team encourages ongoing engagement and collaboration as the project evolves. Key links: Discord: OpenGPT-X Discord server Research Papers: OpenGPT-X News Research Papers LLM Leaderboard: European LLM Leaderboard LLM Leaderboard

Contact Information

You can reach out to the following model card contact:

• OpenGPT-X - contact@opengpt-x.de