Add converters (#27)

analysis/plot_utils.py

@@ -43,7 +43,9 @@ def _compute_category_scores(results: Dict[str, float]) -> Dict[str, float]:
                 {
                     "model": result["subset"].pop("model"),
                     "model_type": result["subset"].pop("model_type"),
-                    "chat_template": result["subset"].pop("chat_template"),
+                    "chat_template": (
+                        result["subset"].pop("chat_template") if "chat_template" in result["subset"] else None
+                    ),
                     # The rewardbench score is the average of the weighted average of the four category scores
                     "score": sum(result["leaderboard"].values()) / len(result["leaderboard"]),
                     "category_scores": result["leaderboard"],

scripts/convert_orig_scores.py

@@ -0,0 +1,63 @@
+from pathlib import Path
+import json
+import requests
+import argparse
+import logging
+
+from rewardbench.utils import calculate_scores_per_section
+from rewardbench.constants import EXAMPLE_COUNTS, SUBSET_MAPPING
+
+logging.basicConfig(level=logging.INFO)
+
+REWARDBENCH_RESULTS_BASEURL = "https://huggingface.co/datasets/allenai/reward-bench-results/resolve/main/eval-set"
+
+
+def get_args():
+    # fmt: off
+    parser = argparse.ArgumentParser(description="Convert RewardBench scores to our formatting")
+    parser.add_argument("--model", required=True, help="JSON URL to convert.")
+    # fmt: on
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    url = f"{REWARDBENCH_RESULTS_BASEURL}/{args.model}.json"
+    response = requests.get(url)
+    if response.ok:
+        data = json.loads(response.text)
+    else:
+        raise ValueError("Cannot parse response")
+
+    results_leaderboard = calculate_scores_per_section(EXAMPLE_COUNTS, SUBSET_MAPPING, data)
+
+    dataset_name = "allenai/reward-bench"
+    if data["model_type"] == "Generative RM":
+        results_dict = {
+            "dataset": dataset_name,
+            "model": data["model"],
+            "chat_template": None if "chat_template" not in data else data["chat_template"],
+            "scores": {},  # can't compute now
+            "leaderboard": results_leaderboard,
+            "subset": data,
+        }
+    else:
+        results_dict = {
+            "accuracy": None,  # can't compute now
+            "num_prompts": 2985,
+            "model": data["model"],
+            "ref_model": None if "ref_model" not in data else data["ref_model"],
+            "tokenizer": None,  # we don't know for now
+            "chat_template": data["chat_template"],
+            "extra_results": data,
+        }
+
+    output_file = f"{data['model'].split('/')[1]}-eng_Latn.json"
+    with open(output_file, "w") as f:
+        json.dump(results_dict, f)
+
+    logging.info(f"Saved to {output_file}")
+
+
+if __name__ == "__main__":
+    main()