IC-1054Code fine-tuning degrades English natural language reasoning in Code LLaMA relative to LLaMA-2, but the effect is negligible or slightly positive in French, Spanish, and German

Suhas Kotha, Jacob Mitchell Springer, Aditi Raghunathan

SourceUnderstanding Catastrophic Forgetting in Language Models via Implicit Inference

The paper evaluates LLaMA-2 and its code-fine-tuned variant Code LLaMA on the XNLI sentence entailment benchmark across four languages. On English, Code LLaMA scores 8.36 percentage points lower than LLaMA-2 (35.90% vs 44.26%). However, on French the difference is -1.45%, on Spanish -0.77%, and on German +1.01%, meaning Code LLaMA performs comparably or slightly better in non-English languages. This suggests the code fine-tuning shifts task inference toward code for English prompts specifically.

Evidence
correlational
Key metric
XNLI accuracy over 2490 test samples: English LLaMA-2 44.26% vs Code LLaMA 35.90% (drop 8.36%); French 33.53% vs 34.98% (-1.45%); Spanish 38.11% vs 38.88% (-0.77%); German 34.50% vs 33.49% (+1.01%)
Caveat
The paper notes that the slight increase in Spanish and French accuracy for Code LLaMA may be due to increased reasoning capabilities associated with code training combined with better task inference, rather than a pure absence of forgetting.
Model
Llama 2 / Llama 2 base, Code Llama
Concepts
Failure mode
Datasets
XNLI [eval], GLUE [source]
Methods
LM Evaluation Harness [eval]
Related findings
IC-1053, IC-1055
Extraction
automatic-extraction