Datasets:

Languages:
English
Size:
n<1K
DOI:
Libraries:
License:
elohn commited on
Commit
980b096
1 Parent(s): 1099902

removed recursive proving to reduce false positives

Browse files
Files changed (1) hide show
  1. baseline.py +3 -0
baseline.py CHANGED
@@ -78,6 +78,9 @@ def benchmark_nextstep(pwd, get_tactics, send_command, search_depth=3, search_wi
78
  for (curr_goal, ps, tac_seq) in old_ps:
79
  next_tactics = get_tactics(curr_goal, prev_lines + '\n'.join(tac_seq))
80
  for next_tactic, _scr in sorted(next_tactics, key=lambda p: -p[1])[:search_width]:
 
 
 
81
  #print('\n'.join(tac_seq + [next_tactic]))
82
  outp, new_proofState = send_tactic(lean_repl, next_tactic, ps)
83
  if outp is None:
 
78
  for (curr_goal, ps, tac_seq) in old_ps:
79
  next_tactics = get_tactics(curr_goal, prev_lines + '\n'.join(tac_seq))
80
  for next_tactic, _scr in sorted(next_tactics, key=lambda p: -p[1])[:search_width]:
81
+ if prop_name in next_tactic:
82
+ continue # although this in theory Can be correct, LEAN DOES NOT CORRECTLY THROW ERRORS when the theorem name is used in a proof.
83
+ # in fact, Lean REPL will return a proofstate with empty goals and no errors! This creates false positives, so we skip these tactics.
84
  #print('\n'.join(tac_seq + [next_tactic]))
85
  outp, new_proofState = send_tactic(lean_repl, next_tactic, ps)
86
  if outp is None: