eduagarcia commited on
Commit
79aba72
1 Parent(s): 1764c92

minor fixes

Browse files
src/leaderboard/read_evals.py CHANGED
@@ -75,6 +75,7 @@ class EvalResult:
75
  tasks = ORIGINAL_TASKS
76
  for task in tasks:
77
  benchmark, metric = task
 
78
 
79
  # We skip old mmlu entries
80
  wrong_mmlu_version = False
@@ -241,7 +242,7 @@ def get_raw_eval_results(results_path: str, requests_path: str, dynamic_path: st
241
  if v.status == "FINISHED" and not v.hidden:
242
  v.to_dict() # we test if the dict version is complete
243
  results.append(v)
244
- except KeyError: # not all eval values present
245
  continue
246
 
247
  return results
 
75
  tasks = ORIGINAL_TASKS
76
  for task in tasks:
77
  benchmark, metric = task
78
+ metric = metric + ',all'
79
 
80
  # We skip old mmlu entries
81
  wrong_mmlu_version = False
 
242
  if v.status == "FINISHED" and not v.hidden:
243
  v.to_dict() # we test if the dict version is complete
244
  results.append(v)
245
+ except KeyError as e: # not all eval values present
246
  continue
247
 
248
  return results
src/scripts/update_all_request_files.py CHANGED
@@ -46,7 +46,8 @@ def update_models(file_path, models, original_leaderboard_files=None):
46
 
47
  # Is the model still on the hub?
48
  model_name = model_id
49
- if model_cfg.card_data is not None and model_cfg.card_data.base_model is not None:
 
50
  model_name = model_cfg.card_data.base_model # for adapters, we look at the parent model
51
  still_on_hub, _, _ = is_model_on_hub(
52
  model_name=model_name, revision=data.get("revision"), trust_remote_code=True, test_tokenizer=False, token=H4_TOKEN
 
46
 
47
  # Is the model still on the hub?
48
  model_name = model_id
49
+
50
+ if model_cfg.card_data is not None and hasattr(model_cfg.card_data, "base_model") and model_cfg.card_data.base_model is not None:
51
  model_name = model_cfg.card_data.base_model # for adapters, we look at the parent model
52
  still_on_hub, _, _ = is_model_on_hub(
53
  model_name=model_name, revision=data.get("revision"), trust_remote_code=True, test_tokenizer=False, token=H4_TOKEN