IC-799WizardMath-70b scores lower than base Llama-2-70b on TabMWP (49.8% vs 57.5%), indicating degraded OOD generalization from rationale-based fine-tuning

Zhibin Gou, Zhihong Shao, Yeyun Gong, yelong shen, Yujiu Yang, Minlie Huang, Nan Duan, Weizhu Chen

SourceToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving

The paper reports that WizardMath-70b, which was fine-tuned with rationale-based mathematical data, underperforms the base Llama-2-70b model on the out-of-distribution TabMWP dataset. WizardMath-70b achieves 49.8% while Llama-2-70b achieves 57.5%. The authors use this as evidence that 'rationale-based fine-tuning negatively affects out-of-distribution generalization.' This is contrasted with TORA-70b, which achieves 74.0% on the same dataset.

Evidence
correlational
Key metric
WizardMath-70b: 49.8% on TabMWP; Llama-2-70b (base): 57.5% on TabMWP
Caveat
The comparison is between a fine-tuned model and its base; the gap may reflect the specific fine-tuning data and procedure of WizardMath rather than a general property of rationale-based training.
Model
WizardMath, Llama 2 / Llama 2 base
Datasets
TabMWP [eval]
Related work
WizardMath [compared-to]
Related findings
IC-798
Extraction
automatic-extraction