up

app.py

@@ -42,7 +42,7 @@ def avg_over_rewardbench(dataframe_core, dataframe_prefs):
     2. Chat Hard: Includes the hard chat subsets (mt-bench-hard, llmbar-natural, llmbar-adver-neighbor, llmbar-adver-GPTInst, llmbar-adver-GPTOut, llmbar-adver-manual)
     3. Safety: Includes the safety subsets (refusals-dangerous, refusals-offensive, xstest-should-refuse, xstest-should-respond, do not answer)
     4. Code: Includes the code subsets (hep-cpp, hep-go, hep-java, hep-js, hep-python, hep-rust)
-    5. Test Sets: Includes the test sets (anthropic_helpful, mtbench_gpt4, shp, summarize)
+    5. Classic Sets: Includes the test sets (anthropic_helpful, mtbench_human, shp, summarize)
     """
     new_df = dataframe_core.copy()
     dataframe_prefs = dataframe_prefs.copy()
@@ -61,28 +61,28 @@ def avg_over_rewardbench(dataframe_core, dataframe_prefs):
     new_df = new_df[keep_columns]
 
     # selected average from pref_sets
-    pref_columns = ["anthropic_helpful", "mtbench_gpt4", "shp", "summarize"]
+    pref_columns = ["anthropic_helpful", "anthropic_hhh", "mtbench_human", "shp", "summarize"]
     pref_data = dataframe_prefs[pref_columns].values
 
     # add column test sets knowing the rows are not identical, take superset
-    dataframe_prefs["Test Sets"] = np.round(np.nanmean(pref_data, axis=1), 2)
+    dataframe_prefs["Classic Sets"] = np.round(np.nanmean(pref_data, axis=1), 2)
 
     # add column Test Sets empty to new_df
-    new_df["Test Sets"] = np.nan
-    # per row in new_df if model is in dataframe_prefs, add the value to new_df["Test Sets"]
+    new_df["Classic Sets"] = np.nan
+    # per row in new_df if model is in dataframe_prefs, add the value to new_df["Classic Sets"]
     values = []
     for i, row in new_df.iterrows():
         model = row["model"]
         if model in dataframe_prefs["model"].values:
-            values.append(dataframe_prefs[dataframe_prefs["model"] == model]["Test Sets"].values[0])
-            # new_df.at[i, "Test Sets"] = dataframe_prefs[dataframe_prefs["model"] == model]["Test Sets"].values[0]
+            values.append(dataframe_prefs[dataframe_prefs["model"] == model]["Classic Sets"].values[0])
+            # new_df.at[i, "Classic Sets"] = dataframe_prefs[dataframe_prefs["model"] == model]["Classic Sets"].values[0]
         else:
             values.append(np.nan)
     
-    new_df["Test Sets"] = values
+    new_df["Classic Sets"] = values
 
     # add total average
-    data_cols += ["Test Sets"]
+    data_cols += ["Classic Sets"]
     new_df["average"] = np.round(np.nanmean(new_df[data_cols].values, axis=1), 2)
 
     # make average third column
@@ -280,7 +280,7 @@ with gr.Blocks(css=custom_css) as app:
         #             elem_id="rewardbench_dataframe_length",
         #             height=1000,
         #         )
-        with gr.TabItem("Existing Test Sets"):
+        with gr.TabItem("Classic Sets"):
             with gr.Row():
                 search_3 = gr.Textbox(label="Model Search (delimit with , )", show_label=False, placeholder="Model Search (delimit with , )")
                 model_types_3 = gr.CheckboxGroup(["Seq. Classifiers", "DPO", "Custom Classifiers", "AI2 Experiments"],
@@ -318,7 +318,8 @@ with gr.Blocks(css=custom_css) as app:
         with gr.TabItem("Dataset Viewer"):
             with gr.Row():
                 # loads one sample
-                gr.Markdown("## Random Dataset Sample Viewer")
+                gr.Markdown("""## Random Dataset Sample Viewer
+Warning, refusals, XSTest, and donotanswer datasets have sensitive content.""")
                 subset_selector = gr.Dropdown(subsets, label="Subset", value=None, multiselect=True)
                 button = gr.Button("Show Random Sample")
 

src/md.py

@@ -9,7 +9,7 @@ We average over 4 core sections (per prompt weighting):
 2. **Chat Hard**: Includes the hard chat subsets (mt-bench-hard, llmbar-natural, llmbar-adver-neighbor, llmbar-adver-GPTInst, llmbar-adver-GPTOut, llmbar-adver-manual)
 3. **Safety**: Includes the safety subsets (refusals-dangerous, refusals-offensive, xstest-should-refuse, xstest-should-respond, do not answer)
 4. **Code**: Includes the code subsets (hep-cpp, hep-go, hep-java, hep-js, hep-python, hep-rust)
-5. **Test Sets**: Includes the test sets (anthropic_helpful, mtbench_gpt4, shp, summarize)
+5. **Classic Sets**: Includes the test sets ([anthropic_helpful](https://huggingface.co/datasets/Anthropic/hh-rlhf), [anthropic_hhh](https://huggingface.co/datasets/HuggingFaceH4/hhh_alignment), [mtbench_human](https://huggingface.co/datasets/lmsys/mt_bench_human_judgments), [shp](https://huggingface.co/datasets/stanfordnlp/SHP), [summarize](https://huggingface.co/datasets/openai/summarize_from_feedback))
 
 We include multiple types of reward models in this evaluation:
 1. **Sequence Classifiers** (Seq. Classifier): A model, normally trained with HuggingFace AutoModelForSequenceClassification, that takes in a prompt and a response and outputs a score.
@@ -82,5 +82,5 @@ For more details, see the [dataset](https://huggingface.co/datasets/ai2-rlhf-col
 TOP_TEXT = """
 # RewardBench: Benchmarking Reward Models
 ### Evaluating the capabilities, safety, and pitfalls of reward models
-[Code](https://github.com/allenai/herm) | [Eval. Dataset](https://huggingface.co/datasets/ai2-adapt-dev/rm-benchmark-dev) | [Existing Test Sets](https://huggingface.co/datasets/allenai/pref-test-sets) | [Results](https://huggingface.co/datasets/ai2-adapt-dev/HERM-Results) | Paper (coming soon)
+[Code](https://github.com/allenai/herm) | [Eval. Dataset](https://huggingface.co/datasets/ai2-adapt-dev/rm-benchmark-dev) | [Classic Test Sets](https://huggingface.co/datasets/allenai/pref-test-sets) | [Results](https://huggingface.co/datasets/ai2-adapt-dev/HERM-Results) | Paper (coming soon)
 """