sigdial_ft_a2

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

This model was used in our experiments in our paper: Elaborative Simplification for German-Language Texts. We have uploaded this model for transparency and replicability of our experiments. If however you are interested in German text simplification in general, we recommend our more recent model.

We fine-tuned meta-llama/Meta-Llama-3-8B-Instruct with a set of ca. 2000 newspaper articles which have been simplified by the Austrian Press Agency. This model was trained with the standard and the A2 level texts.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

• Developed by: Freya Hewett, Hadi Asghari
• Model type: simplification model, text generation
• Language(s) (NLP): German
• License: Apache 2.0
• Finetuned from model: meta-llama/Meta-Llama-3-8B-Instruct

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: GermanElabSimplification
• Paper: Elaborative Simplification for German-Language Texts

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

This model works best for simplifying German-language newspaper articles (news items, not commentaries or editorials). It may work for other types of texts.

Downstream Use

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

We have fine-tuned using only newspaper articles. We have not yet performed extensive out-of-domain testing, but believe that the model's capabilities could be improved by fine-tuning on more diverse data.

<!-- ### Out-of-Scope Use --> <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

As with most text generation models, the model sometimes produces information that is incorrect.

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Please check manually that your output text corresponds to the input text, as factual inconsistencies may have arisen.

How to Get Started with the Model

To load the model using transformers:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
device = "cuda"
tokenizer = AutoTokenizer.from_pretrained("frhew/sigdial_ft_a2")
model = AutoModelForCausalLM.from_pretrained("frhew/sigdial_ft_a2", torch_dtype=torch.float16).to(device)

We used the following prompt at inference to test our model:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
Du bist ein hilfreicher Assistent und hilfst dem User, Texte besser zu verstehen.<|eot_id|><|start_header_id|>user<|end_header_id|>
Kannst du bitte den folgenden Text zusammenfassen und sprachlich auf ein A2-Niveau in Deutsch vereinfachen? Schreibe maximal 5 Sätze.
{input_text}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

A sample of the data used to train our model can be found here.

Training Hyperparameters

<!--- **Training regime:** Our training script can be found [here](https://github.com/fhewett/simba/blob/main/models/train_simba.py). --> <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision --> <!-- #### Speeds, Sizes, Times [optional] --> <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

Evaluation

The right hand side shows the results of the manual evaluation, done on the outputs from each model for 35 texts. M.P. stands for meaning preservation, S for simplification, C for coherence, F for factuality; the score represents the percentage of yes answers. More details on the evaluation can be found in the paper. For all metrics, higher is better.

Summary

Citation

BibTeX:

@inproceedings{hewett-etal-2024-elaborative, title = "Elaborative Simplification for {G}erman-Language Texts", author = "Hewett, Freya and Asghari, Hadi and Stede, Manfred", editor = "Kawahara, Tatsuya and Demberg, Vera and Ultes, Stefan and Inoue, Koji and Mehri, Shikib and Howcroft, David and Komatani, Kazunori", booktitle = "Proceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue", month = sep, year = "2024", address = "Kyoto, Japan", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2024.sigdial-1.3", doi = "10.18653/v1/2024.sigdial-1.3", pages = "29--39"}

APA:

Freya Hewett, Hadi Asghari, and Manfred Stede. 2024. Elaborative Simplification for German-Language Texts. In Proceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 29–39, Kyoto, Japan. Association for Computational Linguistics.

Model Card Contact

frhew