dtcxzyw commited on
Commit
aee328b
Β·
unverified Β·
1 Parent(s): ea982bd
Files changed (1) hide show
  1. app.py +4 -3
app.py CHANGED
@@ -53,14 +53,14 @@ model_cnt = 0
53
  for bug_id, time in bug_id_to_time.items():
54
  timeline_xs.append(time)
55
  timeline_ys.append(0)
56
- timeline_cols.append("Baseline")
57
  LEADERBOARD_DF = get_leaderboard_df(EVAL_REQUESTS_PATH, total_issues)
58
  for row in LEADERBOARD_DF.itertuples():
59
  print(row)
60
  model_cnt += 1
61
  for fix in row.fixed_bug_ids:
62
  timeline_xs.append(bug_id_to_time[fix])
63
- timeline_ys.append(model_cnt)
64
  timeline_cols.append(row.method_id)
65
  timeline_df = pd.DataFrame(
66
  {
@@ -99,7 +99,7 @@ with demo:
99
 
100
  with gr.Tabs(elem_classes="tab-buttons") as tabs:
101
  with gr.TabItem("πŸ… Leaderboard", elem_id="llm-benchmark-tab-table", id=0):
102
- leaderboard = init_leaderboard(LEADERBOARD_DF)
103
  gr.ScatterPlot(
104
  timeline_df,
105
  x="time",
@@ -108,6 +108,7 @@ with demo:
108
  x_label="Time",
109
  y_label="Model",
110
  title="Timeline",
 
111
  )
112
 
113
  with gr.TabItem("πŸš€ Submission", elem_id="llm-benchmark-tab-table", id=1):
 
53
  for bug_id, time in bug_id_to_time.items():
54
  timeline_xs.append(time)
55
  timeline_ys.append(0)
56
+ timeline_cols.append("All")
57
  LEADERBOARD_DF = get_leaderboard_df(EVAL_REQUESTS_PATH, total_issues)
58
  for row in LEADERBOARD_DF.itertuples():
59
  print(row)
60
  model_cnt += 1
61
  for fix in row.fixed_bug_ids:
62
  timeline_xs.append(bug_id_to_time[fix])
63
+ timeline_ys.append(-model_cnt)
64
  timeline_cols.append(row.method_id)
65
  timeline_df = pd.DataFrame(
66
  {
 
99
 
100
  with gr.Tabs(elem_classes="tab-buttons") as tabs:
101
  with gr.TabItem("πŸ… Leaderboard", elem_id="llm-benchmark-tab-table", id=0):
102
+ leaderboard = init_leaderboard(LEADERBOARD_DF[COLS])
103
  gr.ScatterPlot(
104
  timeline_df,
105
  x="time",
 
108
  x_label="Time",
109
  y_label="Model",
110
  title="Timeline",
111
+ y_lim=(-model_cnt - 1, 1),
112
  )
113
 
114
  with gr.TabItem("πŸš€ Submission", elem_id="llm-benchmark-tab-table", id=1):