On Provenance in Model Fusion

• Annie Amorim Universidade Federal Fluminense (UFF)
• João Vitor Moraes Universidade Federal Fluminense (UFF)
• Débora Pina Universidade Federal do Rio de Janeiro (UFRJ)
• Aline Paes Universidade Federal Fluminense (UFF)
• Daniel de Oliveira Universidade Federal Fluminense (UFF)

DOI: https://doi.org/10.5753/sbbd.2026.249455

Resumo

As the number of domain-specific language models grows at a fast pace, managing multiple specialized models becomes a challenge. Model fusion is an alternative that combines models to improve robustness. However, fusion workflows involve steps such as model selection, merging strategies, evaluation, and repeated iterations. Selecting the best combination of source models requires understanding the full data derivation path. In this paper, we present FusionProv, an approach compliant with the W3C PROV recommendation that captures provenance from fusion workflows to support traceability and analytics. We evaluate FusionProv in a hate-speech detection application that leverages fused language models.

Palavras-chave: Provenance, Model Fusion, Language Models, Hate Speech

Referências

Aftan, S. et al. (2023). A survey on bert and its applications. In 2023 20th Learning and Technology Conference (L&T), pages 161–166. IEEE.

Amorim, A. et al. (2025). Exploring language model fusion to improve generalization in portuguese hate speech detection. In 2025 28th FUSION, pages 1–8.

Choshen, L. et al. (2022). Fusing finetuned models for better pretraining. CoRR, abs/2204.03044.

Devlin, J. et al. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proc. of the 2019 NAACL, pages 4171–4186, Minneapolis. ACL.

Freire, J. et al. (2008). Provenance for computational tasks: A survey. Computing in Science & Engineering, 10(3):11–21.

Herschel, M. et al. (2017). A survey on provenance: What for? what form? what from? The VLDB Journal, 26(6):881–906.

Howard, J. et al. (2018). Universal language model fine-tuning for text classification. In Proc. of the 56th Annual Meeting of the ACL, Melbourne, pages 328–339. ACL.

Kerzel, D. et al. (2021). Towards tracking provenance from machine learning notebooks. In International Conference on Knowledge Discovery and Information Retrieval.

Matena, M. et al. (2022). Merging models with fisher-weighted averaging.

Moreau, L. et al. (2015). The rationale of prov. Journal of Web Semantics, 35:235–257.

Padovani, G. et al. (2025). Provenance tracking in large-scale machine learning systems. In Workshop Proceedings of the 54th International Conference on Parallel Processing, ICPP Workshops ’25, page 167–174, New York, NY, USA. Association for Computing Machinery.

Pina, D. et al. (2022). Capturing provenance from deep learning applications using keras-prov and colab: a practical approach. Journal of Information and Data Management, 13(5).

Pina, D. et al. (2025). Breadcrumbs for your deep learning model: Following provenance traces with dlprov. Software Impacts, 23:100730.

Schelter, S. et al. (2017). Automatically tracking metadata and provenance of machine learning experiments.

Schlegel, M. and Sattler, K.-U. (2023). Mlflow2prov: Extracting provenance from machine learning experiments. In Proceedings of the Seventh DEEM. ACM.

Souza, R. et al. (2019). Provenance data in the machine learning lifecycle in computational science and engineering. In 2019 IEEE/ACM Workflows in Support of Large-Scale Science (WORKS), pages 1–10.

Yadav, P. et al. (2023). Ties-merging: Resolving interference when merging models. In Neural Information Processing Systems.