major update

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. Reasoning: Includes the code and math subsets (math-prm, hep-cpp, hep-go, hep-java, hep-js, hep-python, hep-rust)
-    5. Classic Sets: Includes the test sets (anthropic_helpful, mtbench_human, shp, summarize)
+    5. Prior 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", "anthropic_hhh", "mtbench_human", "shp", "summarize"]
+    pref_columns = ["anthropic_helpful", "anthropic_hhh", "shp", "summarize"]
     pref_data = dataframe_prefs[pref_columns].values
 
     # add column test sets knowing the rows are not identical, take superset
-    dataframe_prefs["Classic Sets"] = np.round(np.nanmean(pref_data, axis=1), 2)
+    dataframe_prefs["Prior Sets"] = np.round(np.nanmean(pref_data, axis=1), 2)
 
     # add column Test Sets empty to new_df
-    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"]
+    new_df["Prior Sets"] = np.nan
+    # per row in new_df if model is in dataframe_prefs, add the value to new_df["Prior 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]["Classic Sets"].values[0])
-            # new_df.at[i, "Classic Sets"] = dataframe_prefs[dataframe_prefs["model"] == model]["Classic Sets"].values[0]
+            values.append(dataframe_prefs[dataframe_prefs["model"] == model]["Prior Sets"].values[0])
+            # new_df.at[i, "Prior Sets"] = dataframe_prefs[dataframe_prefs["model"] == model]["Prior Sets"].values[0]
         else:
             values.append(np.nan)
     
-    new_df["Classic Sets"] = values
+    new_df["Prior Sets"] = values
 
     # add total average
-    data_cols += ["Classic Sets"]
+    data_cols += ["Prior Sets"]
     new_df["average"] = np.round(np.nanmean(new_df[data_cols].values, axis=1), 2)
 
     # make average third column
@@ -197,7 +197,11 @@ def regex_table(dataframe, regex, filter_button):
         if "Custom Classifiers" not in filter_button:
             dataframe = dataframe[~dataframe["Model Type"].str.contains("Custom Classifier", case=False, na=False)]
     # Filter the dataframe such that 'model' contains any of the regex patterns
-    return dataframe[dataframe["Model"].str.contains(combined_regex, case=False, na=False)]
+    data = dataframe[dataframe["Model"].str.contains(combined_regex, case=False, na=False)]
+
+    # replace column '' with count/rank
+    data[''] = np.arange(1, 1 + len(data))
+    return data
 
 
 with gr.Blocks(css=custom_css) as app:
@@ -280,7 +284,7 @@ with gr.Blocks(css=custom_css) as app:
         #             elem_id="rewardbench_dataframe_length",
         #             height=1000,
         #         )
-        with gr.TabItem("Classic Sets"):
+        with gr.TabItem("Prior Test 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"],
@@ -291,7 +295,7 @@ with gr.Blocks(css=custom_css) as app:
                                                  )
             with gr.Row():
                 PREF_SET_TEXT = """
-                For more information, see the [dataset](https://huggingface.co/datasets/allenai/pref-test-sets). Only the subsets Anthropic Harmless, Anthropic HHH, MTBench Human, Stanford SHP, and OpenAI's Summarize data are used in the leaderboard.
+                For more information, see the [dataset](https://huggingface.co/datasets/allenai/pref-test-sets). Only the subsets Anthropic Helpful, Anthropic HHH, Stanford SHP, and OpenAI's Summarize data are used in the leaderboard ranking.
                 """
                 gr.Markdown(PREF_SET_TEXT)
             with gr.Row():

src/constants.py

@@ -15,6 +15,7 @@ length_categories = {
     'llmbar-adver-manual': 'False',
     'llmbar-adver-neighbor': 'False',
     'llmbar-natural': 'Neutral',
+    'math-prm': 'Neutral',
     'mt-bench-easy': 'False',
     'mt-bench-hard': 'False',
     'mt-bench-med': 'Neutral',
@@ -31,7 +32,7 @@ example_counts = {
     "mt-bench-easy": 28,
     "mt-bench-med": 40,
     "mt-bench-hard": 37,
-    # "math-prm": 984, # actual length 447, upweighting to be equal to code
+    "math-prm": 984, # actual length 447, upweighting to be equal to code
     "refusals-dangerous": 100,
     "refusals-offensive": 100,
     "llmbar-natural": 100,
@@ -54,6 +55,6 @@ subset_mapping = {
     "Chat": ["alpacaeval-easy", "alpacaeval-length", "alpacaeval-hard", "mt-bench-easy", "mt-bench-med"],
     "Chat Hard": ["mt-bench-hard", "llmbar-natural", "llmbar-adver-neighbor", "llmbar-adver-GPTInst", "llmbar-adver-GPTOut", "llmbar-adver-manual"],
     "Safety": ["refusals-dangerous", "refusals-offensive", "xstest-should-refuse", "xstest-should-respond", "donotanswer"],
-    "Reasoning": [#"math-prm", 
+    "Reasoning": ["math-prm", 
                   "hep-cpp", "hep-go", "hep-java", "hep-js", "hep-python", "hep-rust"]
 }

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. **Reasoning**: Includes the code and math subsets (math-prm, hep-cpp, hep-go, hep-java, hep-js, hep-python, hep-rust)
-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))
+5. **Prior 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.
@@ -20,9 +20,18 @@ We include multiple types of reward models in this evaluation:
 All models are evaluated in fp16 expect for Starling-7B, which is evaluated in fp32.
 Others, such as **Generative Judge** are coming soon.
 
+### Model Types
+
+Currently, we evaluate the following model types:
+1. **Sequence Classifiers**: A model, normally trained with HuggingFace AutoModelForSequenceClassification, that takes in a prompt and a response and outputs a score.
+2. **Custom Classifiers**: Research models with different architectures and training objectives to either take in two inputs at once or generate scores differently (e.g. PairRM and Stanford SteamSHP).
+3. **DPO**: Models trained with Direct Preference Optimization (DPO) with a reference model being either the base or supervised fine-tuning checkpoint.
+
+Support of DPO models without a reference model is coming soon.
+
 ### Subset Details
 
-Total number of the prompts is: 2538, filtered from 4676.
+Total number of the prompts is: 2985, filtered from 5123.
 
 | Subset             | Num. Samples (Pre-filtering, post-filtering) | Description |
 | :---------- | :-----: | :---------: |
@@ -77,11 +86,11 @@ Lengths (mean, std. dev.) include the prompt
 | xstest-should-refuse  | False       | 584 (419)      | 904 (493)        | 129 (89)        | 217 (115)         | 81 (47)                | 116 (53)                 |
 | xstest-should-respond | True        | 771 (420)      | 466 (427)        | 189 (105)       | 107 (94)          | 104 (48)               | 67 (48)                  |
 
-For more details, see the [dataset](https://huggingface.co/datasets/ai2-rlhf-collab/rm-benchmark-dev).
+For more details, see the [dataset](https://huggingface.co/datasets/allenai/reward-bench).
 """
 
 TOP_TEXT = """
-# RewardBench: Benchmarking Reward Models
+# RewardBench: Evaluating 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) | [Classic 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/reward-bench) | [Eval. Dataset](https://huggingface.co/datasets/allenai/reward-bench) | [Classic Test Sets](https://huggingface.co/datasets/allenai/pref-test-sets) | [Results](https://huggingface.co/datasets/allenai/reward-bench-results) | Paper (coming soon)
 """

src/utils.py

@@ -116,4 +116,7 @@ def load_all_data(data_repo, subdir:str, subsubsets=False):    # use HF api to p
         cols.insert(1, cols.pop(cols.index('model_type')))
         df = df.loc[:, cols]
 
+    # remove models with DPO Ref. Free as type (future work)
+    df = df[~df["model_type"].str.contains("DPO Ref. Free", na=False)]
+
     return df