Guardado en:
Detalles Bibliográficos
Autores principales: Pande, Manu, Kumar, Shahil, Damle, Anay Yatin
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2506.18535
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909656861900800
author Pande, Manu
Kumar, Shahil
Damle, Anay Yatin
author_facet Pande, Manu
Kumar, Shahil
Damle, Anay Yatin
contents This paper investigates the counterintuitive phenomenon where fine-tuning pre-trained transformer models degrades performance on the MS MARCO passage ranking task. Through comprehensive experiments involving five model variants-including full parameter fine-tuning and parameter efficient LoRA adaptations-we demonstrate that all fine-tuning approaches underperform the base sentence-transformers/all- MiniLM-L6-v2 model (MRR@10: 0.3026). Our analysis reveals that fine-tuning disrupts the optimal embedding space structure learned during the base model's extensive pre-training on 1 billion sentence pairs, including 9.1 million MS MARCO samples. UMAP visualizations show progressive embedding space flattening, while training dynamics analysis and computational efficiency metrics further support our findings. These results challenge conventional wisdom about transfer learning effectiveness on saturated benchmarks and suggest architectural innovations may be necessary for meaningful improvements.
format Preprint
id arxiv_https___arxiv_org_abs_2506_18535
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle When Fine-Tuning Fails: Lessons from MS MARCO Passage Ranking
Pande, Manu
Kumar, Shahil
Damle, Anay Yatin
Computation and Language
Information Retrieval
This paper investigates the counterintuitive phenomenon where fine-tuning pre-trained transformer models degrades performance on the MS MARCO passage ranking task. Through comprehensive experiments involving five model variants-including full parameter fine-tuning and parameter efficient LoRA adaptations-we demonstrate that all fine-tuning approaches underperform the base sentence-transformers/all- MiniLM-L6-v2 model (MRR@10: 0.3026). Our analysis reveals that fine-tuning disrupts the optimal embedding space structure learned during the base model's extensive pre-training on 1 billion sentence pairs, including 9.1 million MS MARCO samples. UMAP visualizations show progressive embedding space flattening, while training dynamics analysis and computational efficiency metrics further support our findings. These results challenge conventional wisdom about transfer learning effectiveness on saturated benchmarks and suggest architectural innovations may be necessary for meaningful improvements.
title When Fine-Tuning Fails: Lessons from MS MARCO Passage Ranking
topic Computation and Language
Information Retrieval
url https://arxiv.org/abs/2506.18535