Merge branch 'main' of https://huggingface.co/spaces/BramVanroy/open_dutch_llm_leaderboard

app.py

@@ -148,7 +148,7 @@ class ResultSet:
         df = pd.DataFrame(data)
         df = df.sort_values(by=self.column_names["average"], ascending=False)
         number_cols = [col for attr, col in self.column_names.items() if attr in TASK_METRICS or attr == "average"]
-        styler = df.style.format("{:.2f}", subset=number_cols, na_rep="<missing>")
+        styler = df.style.format("{:.4f}", subset=number_cols, na_rep="<missing>")
 
         def highlight_max(col):
             return np.where(col == np.nanmax(col.to_numpy()), "font-weight: bold;", None)

content.py

@@ -18,7 +18,9 @@ I did not verify their translations and I do not maintain the datasets, I only r
 
 DISCLAIMER = """## Disclaimer
 
-I did not verify the (translation) quality of the benchmarks. If you encounter issues with the benchmark contents, please contact the original authors.
+**Evaluating generative models.** Counter-intuitively, we often evaluate generative models with multiple choice questions (as done here). This is useful to gauge the reasoning capabilities of LLMs. However, they do not account for the user experience, including how fluent and natural the text is. A prime example is how top models such as Zephyr, Mistral and Mixtral are actually quite poor when using them as a chatbot for Dutch. But they appear to be good at at least "understanding" a task in Dutch and correctly reasoning about it. Similarly, for humans understanding the general gist of a (new) written language (like after a few months on Duolingo) is something completely different from writing an eloquent, native-level article. This is an under-researched part of evaluating LLMs, especially in non-English languages.
+
+**Translations of benchmarks.** I did not verify the (translation) quality of the benchmarks. If you encounter issues with the benchmark contents, please contact the original authors.
 
 I am aware that benchmarking models on *translated* data is not ideal. However, for Dutch there are no other options for generative models at the moment. Because the benchmarks were automatically translated, some translationese effects may occur: the translations may not be fluent Dutch or still contain artifacts of the source text (like word order, literal translation, certain vocabulary items). Because of that, an unfair advantage may be given to the non-Dutch models: Dutch is closely related to English, so if the benchmarks are in automatically translated Dutch that still has English properties, those English models may not have too many issues with that. If the benchmarks were to have been manually translated or, even better, created from scratch in Dutch, those non-Dutch models may have a harder time. Maybe not. We cannot know for sure until we have high-quality, manually crafted benchmarks for Dutch.