Hugging Face's logo

Spaces:
hallucinations-leaderboard
/
leaderboard
Running on CPU Upgrade

App Files Community
27
pminervini commited on
Commit
d7b202a
•
1 Parent(s): 8e3d8c1

update

Browse files
Files changed (7) hide show
  1. src/backend/tasks/halueval/halueval_dialogue.yaml +1 -1
  2. src/backend/tasks/halueval/halueval_qa.yaml +1 -1
  3. src/backend/tasks/halueval/halueval_summarization.yaml +1 -1
  4. src/backend/tasks/memo-trap/memo-trap.yaml +1 -1
  5. src/backend/tasks/memo-trap/memo-trap_v2.yaml +2 -2
  6. src/backend/tasks/nq8/nq8.yaml +1 -1
  7. src/backend/tasks/tqa8/tqa8.yaml +1 -1
src/backend/tasks/halueval/halueval_dialogue.yaml CHANGED
@@ -26,4 +26,4 @@ metric_list:
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
- - version: 0.0
 
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
+ version: 0.0
src/backend/tasks/halueval/halueval_qa.yaml CHANGED
@@ -26,4 +26,4 @@ metric_list:
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
- - version: 0.0
 
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
+ version: 0.0
src/backend/tasks/halueval/halueval_summarization.yaml CHANGED
@@ -26,4 +26,4 @@ metric_list:
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
- - version: 0.0
 
26
  aggregation: mean
27
  higher_is_better: true
28
  metadata:
29
+ version: 0.0
src/backend/tasks/memo-trap/memo-trap.yaml CHANGED
@@ -16,4 +16,4 @@ metric_list:
16
  aggregation: mean
17
  higher_is_better: true
18
  metadata:
19
- - version: 0.0
 
16
  aggregation: mean
17
  higher_is_better: true
18
  metadata:
19
+ version: 0.0
src/backend/tasks/memo-trap/memo-trap_v2.yaml CHANGED
@@ -5,7 +5,7 @@ output_type: multiple_choice
5
  training_split: null
6
  validation_split: data
7
  test_split: null
8
- num_fewshot: 0
9
  doc_to_text: "{{prompt}}"
10
  doc_to_target: answer_index
11
  doc_to_choice: "{{classes}}"
@@ -17,4 +17,4 @@ metric_list:
17
  aggregation: mean
18
  higher_is_better: true
19
  metadata:
20
- - version: 0.0
 
5
  training_split: null
6
  validation_split: data
7
  test_split: null
8
+ # num_fewshot: 0
9
  doc_to_text: "{{prompt}}"
10
  doc_to_target: answer_index
11
  doc_to_choice: "{{classes}}"
 
17
  aggregation: mean
18
  higher_is_better: true
19
  metadata:
20
+ version: 0.0
src/backend/tasks/nq8/nq8.yaml CHANGED
@@ -29,4 +29,4 @@ metric_list:
29
  regexes_to_ignore:
30
  - "\\b(?:The |the |An |A |The |a |an )"
31
  metadata:
32
- - version: 0.0
 
29
  regexes_to_ignore:
30
  - "\\b(?:The |the |An |A |The |a |an )"
31
  metadata:
32
+ version: 0.0
src/backend/tasks/tqa8/tqa8.yaml CHANGED
@@ -28,4 +28,4 @@ metric_list:
28
  ignore_case: true
29
  ignore_punctuation: true
30
  metadata:
31
- - version: 2.0
 
28
  ignore_case: true
29
  ignore_punctuation: true
30
  metadata:
31
+ version: 2.0