fix: Prompt Lean code extraction

This commit is contained in:
Leni Aniva 2024-10-07 18:58:35 -07:00
parent 30cd3063f9
commit 76eb57b22e
Signed by: aniva
GPG Key ID: 4D9B1C8D10EA4C50
4 changed files with 94 additions and 40 deletions

View File

@ -1,4 +1,4 @@
import sys, os, json, subprocess
import sys, os, json, subprocess, time, datetime
from dataclasses import dataclass, asdict, field
from pathlib import Path
from typing import Union, Any, Tuple, Optional
@ -44,6 +44,7 @@ class DatumResult:
Result from one DSP data point
"""
name: str
duration: float
success: Optional[bool] = False
proves: list[Union[SearchResult, SketchParseFailure]] = field(default_factory=list)
@ -106,8 +107,8 @@ def autoformalize_prob(
""" Autoformalize natural language problem to formal language problem. """
pass
@retry(stop=stop_after_attempt(15), wait=wait_exponential(multiplier=2, max=128))
def draft(
#@retry(stop=stop_after_attempt(15), wait=wait_exponential(multiplier=2, max=128))
def step_draft(
eng: Engine,
datum: Datum,
verbose: bool = False,
@ -140,8 +141,8 @@ def draft(
drafts: list[str] = completions
return drafts
@retry(stop=stop_after_attempt(15), wait=wait_exponential(multiplier=2, max=128))
def sketch(
#@retry(stop=stop_after_attempt(15), wait=wait_exponential(multiplier=2, max=128))
def step_sketch(
eng: Engine,
datum: Datum,
drafts: list[str],
@ -181,7 +182,7 @@ def sketch(
# Return
return sketches, x_fl_problem
def prove(
def step_prove(
eng: Engine,
server: Server,
fl_prob: str,
@ -198,7 +199,7 @@ def prove(
# If this throws index out of bound errors it means the source doesn't contain walled off Lean sections.
print(colored("Sketch:", "yellow"), fl_sketch)
lean_code = "\n".join(extract_lean_code(fl_sketch))
print(colored("Lean code:", "light_grey"), lean_code)
print(colored("Lean code:", "light_grey", attrs=["bold"]), lean_code)
try:
states = server.load_sorry(lean_code)
@ -246,31 +247,37 @@ def single_proof_search_dsp_lean(
server_func,
datum: Datum,
) -> DatumResult:
start_time = time.time()
# -- Draft: [y_nl_pred_draft]_n ~ draft(eng, x_nl_prob, P_draft)
y_nl_pred_drafts = draft(eng, datum)
y_nl_pred_drafts = step_draft(eng, datum)
# -- Sketch: z_fl_pred_sketch ~ sketch(eng, x_nl_prob, [y_nl_pred_draft]_n, x_fl_prob, P_sketch)
z_fl_pred_sketches, x_fl_prob = sketch(eng, datum, y_nl_pred_drafts)
z_fl_pred_sketches, x_fl_prob = step_sketch(eng, datum, y_nl_pred_drafts)
assert len(z_fl_pred_sketches) == eng.sketch_sampling_params.top_p
assert len(z_fl_pred_sketches) == eng.sketch_sampling_params.n
server = server_func()
results = []
success = False
for sketch in z_fl_pred_sketches:
for i_sketch, sketch in enumerate(z_fl_pred_sketches):
if len(z_fl_pred_sketches):
print(colored(f"Sketch {1+i_sketch}/{len(z_fl_pred_sketches)}", attrs=["bold", "underline"]))
# -- Prove: y_fl = prove(eng, x_fl_prob, z_fl_pred_sketches)
prove_result = prove(eng, server, x_fl_prob, sketch)
prove_result = step_prove(eng, server, x_fl_prob, sketch)
results.append(prove_result)
if isinstance(prove_result, SearchResult) and prove_result.success:
success = True
break
duration = time.time() - start_time
return DatumResult(
name=str(datum),
success=success,
proves=results,
duration=duration,
)
def full_proof_search_dsp_lean(
@ -279,28 +286,27 @@ def full_proof_search_dsp_lean(
data: list[Datum],
path_output: Path,
):
print(colored(f"DSP on {len(data)} points", "blue", attrs=["bold", "underline"]))
n_success = 0
n_tried = 0
# -- Proof search by DSP over all eval data
for i, datum in tqdm(enumerate(data), total=len(data), desc='DSP proof loop per data point in benchmark.'):
file_name = path_output / f"{i:03}.json"
output_path = path_output / f"{i:03}.json"
key = str(datum)
# Detect if file exists
if file_name.is_file():
obj = json.load(open(file_name, "r"))
if output_path.is_file():
obj = json.load(open(output_path, "r"))
if obj['name'] != key:
print(colored(f"Existing datum name {obj['name']} does not match dataset {key}. The output directory may be wrong"))
return
print(f"Skipped {i}:", colored(key, "green"))
print(f"Skipped {output_path.name}:", colored(key, "green"))
continue
n_tried += 1
print(f"Problem {i}:", colored(key, "cyan"))
print(f"Problem {output_path.name}:", colored(key, "cyan"))
result = single_proof_search_dsp_lean(eng, server_func, datum)
with open(file_name, 'w') as f:
with open(output_path, 'w') as f:
json.dump(asdict(result), f)
if result.success:
n_success += 1
@ -337,7 +343,6 @@ def load_data(args) -> list[Datum]:
# -- Main
def main(args):
import time, datetime
start_time = time.time()
# Setup paths and data
@ -369,7 +374,7 @@ def main(args):
# - Run DSP for Lean
api_key = os.environ['OPENAI_API_KEY']
draft_sampling_params = SamplingParams(
n=args.n_samples,
n=1, #args.n_samples,
max_tokens=args.max_tokens,
top_p=args.top_p,
temperature=args.temperature,
@ -390,6 +395,9 @@ def main(args):
sketch_sampling_params=sketch_sampling_params,
)
print(colored(f"DSP on {len(data_eval)} points", "blue", attrs=["bold", "underline"]))
print(f"Draft={draft_sampling_params}")
print(f"Sketch={sketch_sampling_params}")
# - Full proof search with DSP
full_proof_search_dsp_lean(eng, server_func, data_eval, path_output)
@ -457,11 +465,31 @@ if __name__ == "__main__":
)
parser.add_argument("--start", default=0)
parser.add_argument("--end", default=sys.maxsize)
parser.add_argument("--batchsize", default=10, help="putnam has 348")
parser.add_argument("--n-samples", default=1, help="num seqs to return for given prompt")
parser.add_argument("--max-tokens", default=2048, help="Maximum number of tokens in one sample")
parser.add_argument("--top-p", default=0.95, help="Sampling top p")
parser.add_argument("--temperature", default=0.8, help="Sampling temperature")
parser.add_argument(
"--batchsize",
default=10, type=int,
help="putnam has 348",
)
parser.add_argument(
"--n-samples",
default=1, type=int,
help="Number of sketch samples for a draft",
)
parser.add_argument(
"--max-tokens",
default=2048, type=int,
help="Maximum number of tokens in one sample",
)
parser.add_argument(
"--top-p",
default=0.95, type=float,
help="Sampling top p via nucleus sampling",
)
parser.add_argument(
"--temperature",
default=0.8, type=float,
help="Sampling temperature",
)
parser.add_argument("--verbose", action='store_true')
args = parser.parse_args()

View File

@ -143,17 +143,21 @@ prompt_sketch_template_lean4_v0 = get_prompt_sketch_template_4_lean_v0()
WALL = "```"
def postprocess_lean(
code,
placeholder: str = TOKEN_PLACEHOLDER,
):
return code.replace("", "Nat").replace(placeholder, "sorry")
def extract_lean_code(
sketch: str,
placeholder: str = TOKEN_PLACEHOLDER,
strip_imports: bool = True) -> list[str]:
lines = sketch.split("\n")
# find backtick markers ```
if WALL not in sketch:
# No walls found. The whole thing must be code
lines = [line for line in lines if not line.startswith("import ")]
return ["\n".join(lines)]
return [postprocess_lean("\n".join(lines))]
lean_codes = []
curr = []
is_walled = False
@ -172,8 +176,7 @@ def extract_lean_code(
if is_walled_lean:
# found wall
code = "\n".join(curr) + "\n"
code = code.replace("", "Nat").replace(placeholder, "sorry")
lean_codes.append(code)
lean_codes.append(postprocess_lean(code))
curr = []
is_walled = False
is_walled_lean = False
@ -201,10 +204,11 @@ class TestPrompts(unittest.TestCase):
self.assertEqual(len(codes), 1)
def test_extract_sketch_no_wall(self):
payload = "example : forall (n: Prop), n -> n := sorry"
payload = f"example : forall (n: Prop), n -> n := {TOKEN_PLACEHOLDER}"
payload1 = f"\nexample : forall (n: Prop), n -> n := sorry"
sketch = f"import Mathlib\n\n{payload}"
codes = extract_lean_code(sketch)
self.assertEqual(codes, ["\n" + payload])
self.assertEqual(codes, [payload1])
if __name__ == '__main__':
unittest.main()

View File

@ -28,3 +28,6 @@ class HammerAgent(Agent):
self.goal_tactic_id_map[key] = i + 1
return self.tactics[i]
def reset(self):
self.goal_tactic_id_map = collections.defaultdict(lambda : 0)

View File

@ -1,4 +1,5 @@
from abc import abstractmethod
import time
from dataclasses import dataclass
from typing import Optional
import collections, unittest
@ -38,6 +39,8 @@ class SearchState:
@dataclass(frozen=True)
class SearchResult:
n_goals_root: int
duration: float
success: bool
steps: int
@ -77,10 +80,14 @@ class Agent:
max_trials_per_goal: int = 5,
verbose: bool = False) -> SearchResult:
"""
Searches using th
Executes proof search on this state
"""
assert server.is_automatic(), "Search must be run in automatic mode"
n_goals_root = len(goal_state.goals)
time_start = time.time()
initial_state = SearchState(
state=goal_state,
parent=None,
@ -88,9 +95,6 @@ class Agent:
priorities=[0.0 for _ in goal_state.goals]
)
search_stack = [initial_state]
"""
Executes proof search on this state
"""
for i_step in range(max_steps):
assert search_stack, "No states in search stack"
@ -101,7 +105,12 @@ class Agent:
assert isinstance(search_state, SearchState)
if search_state.is_solved:
return SearchResult(success=True, steps=i_step)
return SearchResult(
n_goals_root=n_goals_root,
duration=time.time() - time_start,
success=True,
steps=i_step,
)
# Find the unsolved goal with the highest priority
goal_id = search_state.next_goal_id
@ -124,7 +133,12 @@ class Agent:
if verbose:
print("Search stack has been exhausted")
self.reset()
return SearchResult(success=False, steps=i_step)
return SearchResult(
n_goals_root=n_goals_root,
duration=time.time() - time_start,
success=False,
steps=i_step,
)
continue
try:
@ -156,7 +170,12 @@ class Agent:
print("Search iteration limit exhausted")
self.reset()
return SearchResult(success=False, steps=max_steps)
return SearchResult(
n_goals_root=n_goals_root,
duration=time.time() - time_start,
success=False,
steps=max_steps,
)
class DumbAgent(Agent):