fixes

README.md

@@ -2,8 +2,8 @@
 
 This will hold scripts for generating scores and uploading results.
 Two primary scripts need to be created:
-1. `run_c_rm.py`:
-2. `run_dpo_rm.py`:
+1. `run_rm.py`: Run evaluations for reward models.
+2. `run_dpo.py`: Run evaluations for direct preference optimization (DPO) models.
 
 ## Links
 Dataset, space, etc coming soon.

herm/utils.py

@@ -19,7 +19,7 @@
 from fastchat.conversation import Conversation
 from transformers import PreTrainedTokenizer
 
-CORE_EVAL_SET = "ai2-rlhf-collab/rm-benchmark-dev"
+CORE_EVAL_SET = "ai2-adapt-dev/rm-benchmark-dev"
 EXTRA_PREF_SETS = "allenai/pref-test-sets"
 
 

scripts/run_rm.py

@@ -40,8 +40,8 @@
 api = HfApi(token=HF_TOKEN)
 
 # data repo to upload results
-EVAL_REPO = "ai2-rlhf-collab/rm-benchmark-results"
-PREFS_REPO = "ai2-rlhf-collab/rm-testset-results"
+EVAL_REPO = "ai2-adapt-dev/rm-benchmark-results"
+PREFS_REPO = "ai2-adapt-dev/rm-testset-results"
 
 
 def get_args():
@@ -274,6 +274,8 @@ def custom_collate_fn(batch):
     ############################
     # add column for results for easy printing
     out_dataset = dataset.add_column("results", results)
+    # add subsets back (removed so it's not handled by cuda)
+    out_dataset = out_dataset.add_column("subset", subsets)
 
     results = {}
     results["model"] = args.model