Datasets:

Benjamin-eecs
/

openrsi-commit-runtime-assets

Dataset card Files Files and versions

xet

Community

openrsi-commit-runtime-assets / heldout_pool /alvinwan__timefhuman-55d713d4352a /solution /patch.diff

Benjamin-eecs

release v1: 65 held-out tasks, 78 bootstrap records, r2e_local adapter, manifest ae94c40b (part 4)

622387b verified about 1 month ago

Raw History Blame Contribute Delete

21.2 kB

diff --git a/PERFORMANCE.md b/PERFORMANCE.md
index 33e21db..0ec7212 100644
--- a/PERFORMANCE.md
+++ b/PERFORMANCE.md
@@ -4,21 +4,12 @@ Status as of April 2, 2026.
 
 ## Results
 
-- Runtime path: deterministic whole-string parse, extraction-first for noisy text, LALR fallback only on misses. Earley is not used at runtime.
-- Test suite: `127 passed in 0.22s`
-- Tight loop: `19.52 us/input` at `10000` rounds
-- Internal path benchmark:
-  - `fast exact`: `6.5 us`
-  - `fast collection`: `19.5 us`
-  - `extract hit`: `46.1 us`
-  - `extract miss`: `13.6 us`
-- Cross-library snapshot from `benchmarks/benchmark_baselines.py`:
-  - `timefhuman`: `56.1 us/input`, `37/37 ok`, `10/10 exact`
-  - `parsedatetime`: `45.0 us/input`, `36/37 ok`, `6/10 exact`
-  - `datefinder`: `42.9 us/input`, `23/37 ok`, `5/10 exact`
-  - `recurrent`: `195.2 us/input`, `36/37 ok`, `6/10 exact`
-  - `ctparse`: `12808.2 us/input`, `37/37 ok`, `3/10 exact`
-  - `dateparser`: `49105.3 us/input`, `20/37 ok`, `6/10 exact`
+- Runtime path: deterministic whole-string parse first, bounded extraction for noisy text, LALR fallback only on misses. Earley is not used at runtime.
+- Correctness: public `timefhuman(...)` now completes on the `datefinder` `core_corpus` and `seattle_html_76k` datasets without raising conversion errors.
+- Test suite: `137 passed in 0.55s`
+- Warm whole-document median on this machine:
+  - `core_corpus`: `timefhuman 0.00199s` for `10` matches, `datefinder.extract 0.00045s` for `14` matches
+  - `seattle_html_76k`: `timefhuman 0.09717s` for `48` matches, `datefinder.extract 0.07821s` for `57` matches
 
 ## Reproduce
 
@@ -28,36 +19,43 @@ Run tests:
 /tmp/timefhuman-bench-venv/bin/python -m pytest -q
 ```
 
-Run the cross-library benchmark:
-
-```bash
-/tmp/timefhuman-bench-venv/bin/python benchmarks/benchmark_baselines.py
-```
-
-Run the internal path benchmark:
-
-```bash
-/tmp/timefhuman-bench-venv/bin/python benchmarks/benchmark_paths.py
-```
-
-Run the local tight loop:
+Run the warmed whole-document benchmark used above.
+This expects a local clone of `datefinder` at `/tmp/datefinder`.
 
 ```bash
 /tmp/timefhuman-bench-venv/bin/python - <<'PY'
+from datetime import datetime, timezone
+import statistics
 import time
-from datetime import datetime
-from benchmarks.benchmark_baselines import INPUTS
-from timefhuman import timefhuman
-from timefhuman.main import tfhConfig
+from pathlib import Path
 
-cfg = tfhConfig(now=datetime(2018, 8, 4, 14, 0))
-for rounds in [1000, 5000, 10000]:
-    start = time.perf_counter()
-    for _ in range(rounds):
-        for text in INPUTS:
-            timefhuman(text, config=cfg)
-    us = (time.perf_counter() - start) / (rounds * len(INPUTS)) * 1e6
-    print(rounds, f"{us:.2f} us/input")
+import datefinder
+from timefhuman import timefhuman, tfhConfig
+
+ref = datetime(2026, 3, 18, 12, 0, tzinfo=timezone.utc)
+cfg = tfhConfig(now=ref)
+docs = [
+    ("core_corpus", "\n".join(
+        x.strip()
+        for x in Path("/tmp/datefinder/bench/corpus_core.txt").read_text(encoding="utf-8").splitlines()
+        if x.strip()
+    ), 10),
+    ("seattle_html_76k", Path("/tmp/datefinder/tests/seattle_weekly.html").read_text(errors="ignore"), 7),
+]
+
+def bench(fn, iterations):
+    times = []
+    count = 0
+    for _ in range(iterations):
+        start = time.perf_counter()
+        count = fn()
+        times.append(time.perf_counter() - start)
+    return statistics.median(times), count
+
+for name, text, iterations in docs:
+    tfh = bench(lambda: len(timefhuman(text, cfg)), iterations)
+    df = bench(lambda: len(datefinder.extract(text, reference_dt=ref)), iterations)
+    print(name, "timefhuman", tfh, "datefinder", df)
 PY
 ```
 
diff --git a/timefhuman/fastpath.py b/timefhuman/fastpath.py
index 56de861..c2a152a 100644
--- a/timefhuman/fastpath.py
+++ b/timefhuman/fastpath.py
@@ -1,5 +1,5 @@
 import re
-from datetime import timedelta
+from datetime import timedelta, timezone as dt_timezone
 
 from dateutil.relativedelta import relativedelta, weekdays
 import pytz
@@ -11,6 +11,9 @@ from timefhuman.semantics import (
     DATE_TIME_NAME_TO_TEMPLATE,
     MODIFIER_TO_OFFSET,
     NUMBER_WORDS,
+    build_date,
+    build_numeric_date,
+    build_time,
     POSITION_TO_DELTA,
     TIME_NAME_TO_TEMPLATE,
     UNIT_ALIASES,
@@ -78,6 +81,7 @@ NUMBER_UNIT_PATTERN = re.compile(
     r"\s*"
     r"(?P<unit>seconds?|secs?|sec|minutes?|mins?|min|hours?|hour|hrs?|hr|days?|day|weeks?|week|wks?|wk|months?|month|mos|years?|year|yrs?|yr|mo|[smhdwy])"
 )
+NUMERIC_TIMEZONE_OFFSET_PATTERN = re.compile(r"^(?P<body>.*\S)\s+(?P<sign>[+-])(?P<hour>\d{2})(?::?(?P<minute>\d{2}))$")
 
 
 def parse_fast(text: str, config: tfhConfig, timezone_mapping, start_pos: int = 0):
@@ -253,18 +257,22 @@ def _parse_atomic_datetime(text: str, config: tfhConfig, tzinfo):
     lower_body = text.lower()
     iso_match = ISO_PATTERN.fullmatch(text)
     if iso_match:
+        date = build_date(
+            year=int(iso_match.group("year")),
+            month=int(iso_match.group("month")),
+            day=int(iso_match.group("day")),
+        )
+        time = build_time(
+            hour=int(iso_match.group("hour")),
+            minute=int(iso_match.group("minute")),
+            second=int(iso_match.group("second") or 0),
+            millisecond=int(iso_match.group("millisecond") or 0),
+        )
+        if date is None or time is None:
+            return None
         return tfhDatetime(
-            date=tfhDate(
-                year=int(iso_match.group("year")),
-                month=int(iso_match.group("month")),
-                day=int(iso_match.group("day")),
-            ),
-            time=tfhTime(
-                hour=int(iso_match.group("hour")),
-                minute=int(iso_match.group("minute")),
-                second=int(iso_match.group("second") or 0),
-                millisecond=int(iso_match.group("millisecond") or 0),
-            ),
+            date=date,
+            time=time,
             tz=tzinfo,
         )
 
@@ -336,7 +344,7 @@ def _parse_date(text: str, config: tfhConfig):
 
     day_suffix = DAY_SUFFIX_PATTERN.fullmatch(lower)
     if day_suffix:
-        return tfhDate(day=int(day_suffix.group("day")))
+        return build_date(day=int(day_suffix.group("day")))
 
     weekday = _parse_weekday_name(lower)
     if weekday is not None:
@@ -354,15 +362,7 @@ def _parse_numeric_date(text: str):
     first = int(match.group("a"))
     second = int(match.group("b"))
     third = match.group("c")
-
-    if third is None:
-        if second > 31:
-            return tfhDate(month=first, year=normalize_year(second))
-        return tfhDate(month=first, day=second)
-
-    if len(match.group("a")) == 4:
-        return tfhDate(year=first, month=second, day=int(third))
-    return tfhDate(month=first, day=second, year=normalize_year(int(third)))
+    return build_numeric_date(first, second, int(third) if third is not None else None)
 
 
 def _parse_monthname_date(text: str):
@@ -379,10 +379,10 @@ def _parse_monthname_date(text: str):
     year = match.group("year")
 
     if year is not None:
-        return tfhDate(month=month, day=first, year=normalize_year(int(year)))
+        return build_date(month=month, day=first, year=normalize_year(int(year)))
     if suffix or first <= 31:
-        return tfhDate(month=month, day=first)
-    return tfhDate(month=month, year=normalize_year(first))
+        return build_date(month=month, day=first)
+    return build_date(month=month, year=normalize_year(first))
 
 
 def _parse_day_month_or_year(text: str):
@@ -393,12 +393,12 @@ def _parse_day_month_or_year(text: str):
             return None
         year = match.group("year")
         if year is not None:
-            return tfhDate(day=int(match.group("day")), month=month, year=normalize_year(int(year)))
-        return tfhDate(day=int(match.group("day")), month=month)
+            return build_date(day=int(match.group("day")), month=month, year=normalize_year(int(year)))
+        return build_date(day=int(match.group("day")), month=month)
 
     match = DAY_YEAR_PATTERN.fullmatch(text)
     if match:
-        return tfhDate(day=int(match.group("day")), year=normalize_year(int(match.group("year"))))
+        return build_date(day=int(match.group("day")), year=normalize_year(int(match.group("year"))))
 
     return None
 
@@ -423,7 +423,7 @@ def _parse_time_component(text: str, allow_houronly: bool):
     if meridiem is None and match.group("minute") is None and not allow_houronly:
         return None
 
-    return tfhTime(
+    return build_time(
         hour=int(match.group("hour")),
         minute=int(match.group("minute") or 0),
         second=int(match.group("second") or 0),
@@ -503,6 +503,15 @@ def _consume_word_duration(segment: str):
 
 
 def _strip_trailing_timezone(text: str, timezone_mapping):
+    offset_match = NUMERIC_TIMEZONE_OFFSET_PATTERN.fullmatch(text)
+    if offset_match:
+        hours = int(offset_match.group("hour"))
+        minutes = int(offset_match.group("minute"))
+        if hours <= 23 and minutes <= 59:
+            sign = -1 if offset_match.group("sign") == "-" else 1
+            offset = sign * timedelta(hours=hours, minutes=minutes)
+            return offset_match.group("body").strip(), dt_timezone(offset)
+
     if not text or not text[-1].isalpha():
         return text, None
 
@@ -534,7 +543,7 @@ def _parse_modifier_prefix(tokens):
 
 def _strip_leading_weekday(text: str):
     parts = text.split(maxsplit=1)
-    if len(parts) == 2 and weekday_index(parts[0]) is not None:
+    if len(parts) == 2 and weekday_index(parts[0].rstrip(",")) is not None:
         return parts[1]
     return text
 
diff --git a/timefhuman/grammar.lark b/timefhuman/grammar.lark
index e46bc8d..b55c24f 100644
--- a/timefhuman/grammar.lark
+++ b/timefhuman/grammar.lark
@@ -24,6 +24,7 @@ FLOAT_NUMBER: /((\d+\.\d*|\.\d+)(e[-+]?\d+)?|\d+(e[-+]?\d+))/
 NUMERIC_DATE.2: /\d{1,4}\/\d{1,4}(?:\/\d{1,4})?|\d{1,4}\-\d{1,4}\-\d{1,4}/
 ISO_DATETIME.3: /\d{4}-\d{1,2}-\d{1,2}T\d{1,2}:\d{2}(?::\d{2}(?:\.\d{1,6})?)?/
 TIMEZONE: /(?<![a-z])((TIMEZONE_MAPPING))(?![a-z])/
+TIMEZONE_OFFSET.2: /[+-]\d{2}:?\d{2}/
 MODIFIER: /(?<![a-z])(next|last|this|upcoming|previous|past)(?![a-z])/
 POSITION: /(?<![a-z])(first|second|third|fourth|last)(?![a-z])/
 
@@ -97,7 +98,7 @@ modified_weekday: modifier WEEKDAY
                 | modifier modifier WEEKDAY
 modified_month: modifier MONTHNAME
               | modifier modifier MONTHNAME
-weekday: WEEKDAY
+weekday: WEEKDAY (",")?
 monthname: MONTHNAME
 datename: DATENAME
 dayoryear: INT
@@ -114,6 +115,6 @@ second: INT
 millisecond: INT
 meridiem: MERIDIEM
 houronly: INT
-timezone: TIMEZONE
+timezone: TIMEZONE | TIMEZONE_OFFSET
 modifier: MODIFIER
 position: POSITION
diff --git a/timefhuman/lalr.py b/timefhuman/lalr.py
index 4b56fc3..eff8795 100644
--- a/timefhuman/lalr.py
+++ b/timefhuman/lalr.py
@@ -1,9 +1,10 @@
-from datetime import timedelta
+from datetime import timedelta, timezone as dt_timezone
 from pathlib import Path
 import re
 
 from dataclasses import replace
 from lark import Lark, Transformer, UnexpectedInput
+from lark.exceptions import VisitError
 import pytz
 from dateutil.relativedelta import relativedelta, weekdays
 
@@ -14,6 +15,9 @@ from timefhuman.semantics import (
     DATE_TIME_NAME_TO_TEMPLATE,
     MODIFIER_TO_OFFSET,
     NUMBER_WORDS,
+    build_date,
+    build_numeric_date,
+    build_time,
     POSITION_TO_DELTA,
     TIME_NAME_TO_TEMPLATE,
     UNIT_ALIASES,
@@ -85,7 +89,12 @@ def parse_lalr_renderers(string: str, config: tfhConfig, start_pos: int = 0):
     except UnexpectedInput:
         return None
 
-    renderers = tfhTransformer(config=config).transform(tree)
+    try:
+        renderers = tfhTransformer(config=config).transform(tree)
+    except VisitError as exc:
+        if isinstance(exc.orig_exc, ValueError):
+            return None
+        raise
     for renderer in renderers:
         renderer.matched_text_pos = (span_start, span_end)
     return renderers
@@ -186,16 +195,10 @@ class tfhTransformer(Transformer):
         sep = '/' if '/' in value else '-'
         parts = [int(part) for part in value.split(sep)]
         first, second = parts[0], parts[1]
-
-        if len(parts) == 2:
-            if second > 31:
-                return {'date': tfhDate(month=first, year=normalize_year(second))}
-            return {'date': tfhDate(month=first, day=second)}
-
-        third = parts[2]
-        if first >= 1000:
-            return {'date': tfhDate(year=first, month=second, day=third)}
-        return {'date': tfhDate(month=first, day=second, year=normalize_year(third))}
+        result = build_numeric_date(first, second, parts[2] if len(parts) == 3 else None)
+        if result is None:
+            raise ValueError(f"Invalid numeric date: {value}")
+        return {'date': result}
 
     def day(self, children):
         return {'day': int(children[0].value)}
@@ -219,16 +222,28 @@ class tfhTransformer(Transformer):
         year = match.group('year')
 
         if year is not None:
-            return {'date': tfhDate(month=month, day=first, year=normalize_year(int(year)))}
+            result = build_date(month=month, day=first, year=normalize_year(int(year)))
+            if result is None:
+                raise ValueError(f"Invalid monthname date: {children[0].value}")
+            return {'date': result}
         if match.group('suffix') or first <= 31:
-            return {'date': tfhDate(month=month, day=first)}
-        return {'date': tfhDate(month=month, year=normalize_year(first))}
+            result = build_date(month=month, day=first)
+            if result is None:
+                raise ValueError(f"Invalid monthname date: {children[0].value}")
+            return {'date': result}
+        result = build_date(month=month, year=normalize_year(first))
+        if result is None:
+            raise ValueError(f"Invalid monthname date: {children[0].value}")
+        return {'date': result}
 
     def day_ordinal(self, children):
         match = DAY_ORDINAL_PATTERN.fullmatch(children[0].value)
         if match is None:
             raise NotImplementedError(f"Unknown day ordinal: {children[0].value}")
-        return {'date': tfhDate(day=int(match.group('day')))}
+        result = build_date(day=int(match.group('day')))
+        if result is None:
+            raise ValueError(f"Invalid day ordinal: {children[0].value}")
+        return {'date': result}
 
     def modified_month(self, children):
         offset = sum(child['offset'] for child in children[:-1])
@@ -274,13 +289,16 @@ class tfhTransformer(Transformer):
         if 'time' in data:
             return {'time': data['time']}
 
-        return {'time': tfhTime(
+        result = build_time(
             hour=int(data.get("hour", 0)),
             minute=int(data.get("minute", 0)),
             second=int(data.get("second", 0)),
             millisecond=int(data.get("millisecond", 0)),
-            meridiem=data.get("meridiem", None)
-        )}
+            meridiem=data.get("meridiem", None),
+        )
+        if result is None:
+            raise ValueError(f"Invalid time: {data}")
+        return {'time': result}
 
     def meridiem(self, children):
         meridiem = children[0].value.lower()
@@ -292,6 +310,14 @@ class tfhTransformer(Transformer):
 
     def timezone(self, children):
         timezone = children[0].value.lower()
+        if timezone.startswith(("+", "-")):
+            sign = -1 if timezone[0] == "-" else 1
+            body = timezone[1:].replace(":", "")
+            hours = int(body[:2])
+            minutes = int(body[2:])
+            if hours > 23 or minutes > 59:
+                raise ValueError(f"Invalid timezone offset: {timezone}")
+            return {'timezone': dt_timezone(sign * timedelta(hours=hours, minutes=minutes))}
         return {'timezone': pytz.timezone(timezone_mapping[timezone])}
 
     def timename(self, children):
@@ -301,7 +327,10 @@ class tfhTransformer(Transformer):
         return {'time': clone_time(TIME_NAME_TO_TEMPLATE[timename])}
 
     def houronly(self, children):
-        return {'time': tfhTime(hour=int(children[0].value))}
+        result = build_time(hour=int(children[0].value))
+        if result is None:
+            raise ValueError(f"Invalid hour: {children[0].value}")
+        return {'time': result}
 
     def datetimename(self, children):
         datetimename = children[0].value.lower()
@@ -326,7 +355,11 @@ class tfhTransformer(Transformer):
         if len(time_parts) > 2:
             second = time_parts[2]
 
+        date = build_date(year=year, month=month, day=day)
+        time = build_time(hour=hour, minute=minute, second=second, millisecond=millisecond)
+        if date is None or time is None:
+            raise ValueError(f"Invalid ISO datetime: {value}")
         return {'datetime': tfhDatetime(
-            date=tfhDate(year=year, month=month, day=day),
-            time=tfhTime(hour=hour, minute=minute, second=second, millisecond=millisecond),
+            date=date,
+            time=time,
         )}
diff --git a/timefhuman/main.py b/timefhuman/main.py
index 784f72e..d285e2f 100644
--- a/timefhuman/main.py
+++ b/timefhuman/main.py
@@ -57,10 +57,25 @@ def _parse_renderers(string: str, config: tfhConfig):
 
 
 def _matched_results(string: str, renderers, config: tfhConfig):
-    positions = [(renderer.matched_text_pos[0], renderer.matched_text_pos[1]) for renderer in renderers]
-    matched_texts = [string[start:end] for start, end in positions]
-    datetimes = [renderer.to_object(config) for renderer in renderers]
-    return list(zip(matched_texts, positions, datetimes))
+    results = []
+    for renderer in renderers:
+        try:
+            value = renderer.to_object(config)
+        except ValueError:
+            continue
+        start, end = renderer.matched_text_pos
+        results.append((string[start:end], (start, end), value))
+    return results
+
+
+def _valid_objects(renderers, config: tfhConfig):
+    results = []
+    for renderer in renderers:
+        try:
+            results.append(renderer.to_object(config))
+        except ValueError:
+            continue
+    return results
 
 
 def build_raw_tree(string: str, config: tfhConfig):
@@ -97,4 +112,4 @@ def timefhuman(string, config: tfhConfig = DEFAULT_CONFIG, raw: bool=False, now:
 
     if config.return_matched_text:
         return _matched_results(string, renderers, config)
-    return [renderer.to_object(config) for renderer in renderers]
+    return _valid_objects(renderers, config)
diff --git a/timefhuman/semantics.py b/timefhuman/semantics.py
index bfc6317..cbb5d9e 100644
--- a/timefhuman/semantics.py
+++ b/timefhuman/semantics.py
@@ -1,8 +1,8 @@
-from datetime import timedelta
+from datetime import date, timedelta
 
 from dateutil.relativedelta import relativedelta
 
-from timefhuman.renderers import tfhDatetime, tfhTime
+from timefhuman.renderers import tfhDate, tfhDatetime, tfhTime
 from timefhuman.utils import get_month_mapping
 
 
@@ -103,6 +103,69 @@ def normalize_year(value: int):
     return value
 
 
+def build_date(year: int | None = None, month: int | None = None, day: int | None = None):
+    if year is not None and not 1 <= year <= 9999:
+        return None
+    if month is not None and not 1 <= month <= 12:
+        return None
+    if day is not None and not 1 <= day <= 31:
+        return None
+    if month is not None and day is not None:
+        validation_year = year if year is not None else 2000
+        try:
+            date(validation_year, month, day)
+        except ValueError:
+            return None
+    return tfhDate(year=year, month=month, day=day)
+
+
+def build_numeric_date(first: int, second: int, third: int | None = None):
+    if third is None:
+        if first >= 1000:
+            return build_date(year=first, month=second)
+        if second > 31:
+            return build_date(month=first, year=normalize_year(second))
+        if 1 <= first <= 12:
+            return build_date(month=first, day=second)
+        if 1 <= second <= 12:
+            return build_date(month=second, day=first)
+        return None
+
+    if first >= 1000:
+        return build_date(year=first, month=second, day=third)
+
+    year = normalize_year(third)
+    if 1 <= first <= 12:
+        return build_date(month=first, day=second, year=year)
+    if 1 <= second <= 12:
+        return build_date(year=year, month=second, day=first)
+    return None
+
+
+def build_time(
+    hour: int,
+    minute: int = 0,
+    second: int = 0,
+    millisecond: int = 0,
+    meridiem=None,
+):
+    if not 0 <= minute <= 59 or not 0 <= second <= 59 or not 0 <= millisecond <= 999999:
+        return None
+    if meridiem is None:
+        if not 0 <= hour <= 23:
+            return None
+    elif not 1 <= hour <= 12:
+        return None
+
+    return tfhTime(
+        hour=hour,
+        minute=minute,
+        second=second,
+        millisecond=millisecond,
+        meridiem=meridiem,
+    )
+
+
 def clone_time(template: tfhTime):
     return tfhTime(
         hour=template.hour,

	diff --git a/PERFORMANCE.md b/PERFORMANCE.md
	index 33e21db..0ec7212 100644
	--- a/PERFORMANCE.md
	+++ b/PERFORMANCE.md
	@@ -4,21 +4,12 @@ Status as of April 2, 2026.

	## Results

	-- Runtime path: deterministic whole-string parse, extraction-first for noisy text, LALR fallback only on misses. Earley is not used at runtime.
	-- Test suite: `127 passed in 0.22s`
	-- Tight loop: `19.52 us/input` at `10000` rounds
	-- Internal path benchmark:
	- - `fast exact`: `6.5 us`
	- - `fast collection`: `19.5 us`
	- - `extract hit`: `46.1 us`
	- - `extract miss`: `13.6 us`
	-- Cross-library snapshot from `benchmarks/benchmark_baselines.py`:
	- - `timefhuman`: `56.1 us/input`, `37/37 ok`, `10/10 exact`
	- - `parsedatetime`: `45.0 us/input`, `36/37 ok`, `6/10 exact`
	- - `datefinder`: `42.9 us/input`, `23/37 ok`, `5/10 exact`
	- - `recurrent`: `195.2 us/input`, `36/37 ok`, `6/10 exact`
	- - `ctparse`: `12808.2 us/input`, `37/37 ok`, `3/10 exact`
	- - `dateparser`: `49105.3 us/input`, `20/37 ok`, `6/10 exact`
	+- Runtime path: deterministic whole-string parse first, bounded extraction for noisy text, LALR fallback only on misses. Earley is not used at runtime.
	+- Correctness: public `timefhuman(...)` now completes on the `datefinder` `core_corpus` and `seattle_html_76k` datasets without raising conversion errors.
	+- Test suite: `137 passed in 0.55s`
	+- Warm whole-document median on this machine:
	+ - `core_corpus`: `timefhuman 0.00199s` for `10` matches, `datefinder.extract 0.00045s` for `14` matches
	+ - `seattle_html_76k`: `timefhuman 0.09717s` for `48` matches, `datefinder.extract 0.07821s` for `57` matches

	## Reproduce

	@@ -28,36 +19,43 @@ Run tests:
	/tmp/timefhuman-bench-venv/bin/python -m pytest -q
	```

	-Run the cross-library benchmark:
	-
	-```bash
	-/tmp/timefhuman-bench-venv/bin/python benchmarks/benchmark_baselines.py
	-```
	-
	-Run the internal path benchmark:
	-
	-```bash
	-/tmp/timefhuman-bench-venv/bin/python benchmarks/benchmark_paths.py
	-```
	-
	-Run the local tight loop:
	+Run the warmed whole-document benchmark used above.
	+This expects a local clone of `datefinder` at `/tmp/datefinder`.

	```bash
	/tmp/timefhuman-bench-venv/bin/python - <<'PY'
	+from datetime import datetime, timezone
	+import statistics
	import time
	-from datetime import datetime
	-from benchmarks.benchmark_baselines import INPUTS
	-from timefhuman import timefhuman
	-from timefhuman.main import tfhConfig
	+from pathlib import Path

	-cfg = tfhConfig(now=datetime(2018, 8, 4, 14, 0))
	-for rounds in [1000, 5000, 10000]:
	- start = time.perf_counter()
	- for _ in range(rounds):
	- for text in INPUTS:
	- timefhuman(text, config=cfg)
	- us = (time.perf_counter() - start) / (rounds * len(INPUTS)) * 1e6
	- print(rounds, f"{us:.2f} us/input")
	+import datefinder
	+from timefhuman import timefhuman, tfhConfig
	+
	+ref = datetime(2026, 3, 18, 12, 0, tzinfo=timezone.utc)
	+cfg = tfhConfig(now=ref)
	+docs = [
	+ ("core_corpus", "\n".join(
	+ x.strip()
	+ for x in Path("/tmp/datefinder/bench/corpus_core.txt").read_text(encoding="utf-8").splitlines()
	+ if x.strip()
	+ ), 10),
	+ ("seattle_html_76k", Path("/tmp/datefinder/tests/seattle_weekly.html").read_text(errors="ignore"), 7),
	+]
	+
	+def bench(fn, iterations):
	+ times = []
	+ count = 0
	+ for _ in range(iterations):
	+ start = time.perf_counter()
	+ count = fn()
	+ times.append(time.perf_counter() - start)
	+ return statistics.median(times), count
	+
	+for name, text, iterations in docs:
	+ tfh = bench(lambda: len(timefhuman(text, cfg)), iterations)
	+ df = bench(lambda: len(datefinder.extract(text, reference_dt=ref)), iterations)
	+ print(name, "timefhuman", tfh, "datefinder", df)
	PY
	```

	diff --git a/timefhuman/fastpath.py b/timefhuman/fastpath.py
	index 56de861..c2a152a 100644
	--- a/timefhuman/fastpath.py
	+++ b/timefhuman/fastpath.py
	@@ -1,5 +1,5 @@
	import re
	-from datetime import timedelta
	+from datetime import timedelta, timezone as dt_timezone

	from dateutil.relativedelta import relativedelta, weekdays
	import pytz
	@@ -11,6 +11,9 @@ from timefhuman.semantics import (
	DATE_TIME_NAME_TO_TEMPLATE,
	MODIFIER_TO_OFFSET,
	NUMBER_WORDS,
	+ build_date,
	+ build_numeric_date,
	+ build_time,
	POSITION_TO_DELTA,
	TIME_NAME_TO_TEMPLATE,
	UNIT_ALIASES,
	@@ -78,6 +81,7 @@ NUMBER_UNIT_PATTERN = re.compile(
	r"\s*"
	r"(?P<unit>seconds?\|secs?\|sec\|minutes?\|mins?\|min\|hours?\|hour\|hrs?\|hr\|days?\|day\|weeks?\|week\|wks?\|wk\|months?\|month\|mos\|years?\|year\|yrs?\|yr\|mo\|[smhdwy])"
	)
	+NUMERIC_TIMEZONE_OFFSET_PATTERN = re.compile(r"^(?P<body>.*\S)\s+(?P<sign>[+-])(?P<hour>\d{2})(?::?(?P<minute>\d{2}))$")


	def parse_fast(text: str, config: tfhConfig, timezone_mapping, start_pos: int = 0):
	@@ -253,18 +257,22 @@ def _parse_atomic_datetime(text: str, config: tfhConfig, tzinfo):
	lower_body = text.lower()
	iso_match = ISO_PATTERN.fullmatch(text)
	if iso_match:
	+ date = build_date(
	+ year=int(iso_match.group("year")),
	+ month=int(iso_match.group("month")),
	+ day=int(iso_match.group("day")),
	+ )
	+ time = build_time(
	+ hour=int(iso_match.group("hour")),
	+ minute=int(iso_match.group("minute")),
	+ second=int(iso_match.group("second") or 0),
	+ millisecond=int(iso_match.group("millisecond") or 0),
	+ )
	+ if date is None or time is None:
	+ return None
	return tfhDatetime(
	- date=tfhDate(
	- year=int(iso_match.group("year")),
	- month=int(iso_match.group("month")),
	- day=int(iso_match.group("day")),
	- ),
	- time=tfhTime(
	- hour=int(iso_match.group("hour")),
	- minute=int(iso_match.group("minute")),
	- second=int(iso_match.group("second") or 0),
	- millisecond=int(iso_match.group("millisecond") or 0),
	- ),
	+ date=date,
	+ time=time,
	tz=tzinfo,
	)

	@@ -336,7 +344,7 @@ def _parse_date(text: str, config: tfhConfig):

	day_suffix = DAY_SUFFIX_PATTERN.fullmatch(lower)
	if day_suffix:
	- return tfhDate(day=int(day_suffix.group("day")))
	+ return build_date(day=int(day_suffix.group("day")))

	weekday = _parse_weekday_name(lower)
	if weekday is not None:
	@@ -354,15 +362,7 @@ def _parse_numeric_date(text: str):
	first = int(match.group("a"))
	second = int(match.group("b"))
	third = match.group("c")
	-
	- if third is None:
	- if second > 31:
	- return tfhDate(month=first, year=normalize_year(second))
	- return tfhDate(month=first, day=second)
	-
	- if len(match.group("a")) == 4:
	- return tfhDate(year=first, month=second, day=int(third))
	- return tfhDate(month=first, day=second, year=normalize_year(int(third)))
	+ return build_numeric_date(first, second, int(third) if third is not None else None)


	def _parse_monthname_date(text: str):
	@@ -379,10 +379,10 @@ def _parse_monthname_date(text: str):
	year = match.group("year")

	if year is not None:
	- return tfhDate(month=month, day=first, year=normalize_year(int(year)))
	+ return build_date(month=month, day=first, year=normalize_year(int(year)))
	if suffix or first <= 31:
	- return tfhDate(month=month, day=first)
	- return tfhDate(month=month, year=normalize_year(first))
	+ return build_date(month=month, day=first)
	+ return build_date(month=month, year=normalize_year(first))


	def _parse_day_month_or_year(text: str):
	@@ -393,12 +393,12 @@ def _parse_day_month_or_year(text: str):
	return None
	year = match.group("year")
	if year is not None:
	- return tfhDate(day=int(match.group("day")), month=month, year=normalize_year(int(year)))
	- return tfhDate(day=int(match.group("day")), month=month)
	+ return build_date(day=int(match.group("day")), month=month, year=normalize_year(int(year)))
	+ return build_date(day=int(match.group("day")), month=month)

	match = DAY_YEAR_PATTERN.fullmatch(text)
	if match:
	- return tfhDate(day=int(match.group("day")), year=normalize_year(int(match.group("year"))))
	+ return build_date(day=int(match.group("day")), year=normalize_year(int(match.group("year"))))

	return None

	@@ -423,7 +423,7 @@ def _parse_time_component(text: str, allow_houronly: bool):
	if meridiem is None and match.group("minute") is None and not allow_houronly:
	return None

	- return tfhTime(
	+ return build_time(
	hour=int(match.group("hour")),
	minute=int(match.group("minute") or 0),
	second=int(match.group("second") or 0),
	@@ -503,6 +503,15 @@ def _consume_word_duration(segment: str):


	def _strip_trailing_timezone(text: str, timezone_mapping):
	+ offset_match = NUMERIC_TIMEZONE_OFFSET_PATTERN.fullmatch(text)
	+ if offset_match:
	+ hours = int(offset_match.group("hour"))
	+ minutes = int(offset_match.group("minute"))
	+ if hours <= 23 and minutes <= 59:
	+ sign = -1 if offset_match.group("sign") == "-" else 1
	+ offset = sign * timedelta(hours=hours, minutes=minutes)
	+ return offset_match.group("body").strip(), dt_timezone(offset)
	+
	if not text or not text[-1].isalpha():
	return text, None

	@@ -534,7 +543,7 @@ def _parse_modifier_prefix(tokens):

	def _strip_leading_weekday(text: str):
	parts = text.split(maxsplit=1)
	- if len(parts) == 2 and weekday_index(parts[0]) is not None:
	+ if len(parts) == 2 and weekday_index(parts[0].rstrip(",")) is not None:
	return parts[1]
	return text

	diff --git a/timefhuman/grammar.lark b/timefhuman/grammar.lark
	index e46bc8d..b55c24f 100644
	--- a/timefhuman/grammar.lark
	+++ b/timefhuman/grammar.lark
	@@ -24,6 +24,7 @@ FLOAT_NUMBER: /((\d+\.\d*\|\.\d+)(e[-+]?\d+)?\|\d+(e[-+]?\d+))/
	NUMERIC_DATE.2: /\d{1,4}\/\d{1,4}(?:\/\d{1,4})?\|\d{1,4}\-\d{1,4}\-\d{1,4}/
	ISO_DATETIME.3: /\d{4}-\d{1,2}-\d{1,2}T\d{1,2}:\d{2}(?::\d{2}(?:\.\d{1,6})?)?/
	TIMEZONE: /(?<![a-z])((TIMEZONE_MAPPING))(?![a-z])/
	+TIMEZONE_OFFSET.2: /[+-]\d{2}:?\d{2}/
	MODIFIER: /(?<![a-z])(next\|last\|this\|upcoming\|previous\|past)(?![a-z])/
	POSITION: /(?<![a-z])(first\|second\|third\|fourth\|last)(?![a-z])/

	@@ -97,7 +98,7 @@ modified_weekday: modifier WEEKDAY
	\| modifier modifier WEEKDAY
	modified_month: modifier MONTHNAME
	\| modifier modifier MONTHNAME
	-weekday: WEEKDAY
	+weekday: WEEKDAY (",")?
	monthname: MONTHNAME
	datename: DATENAME
	dayoryear: INT
	@@ -114,6 +115,6 @@ second: INT
	millisecond: INT
	meridiem: MERIDIEM
	houronly: INT
	-timezone: TIMEZONE
	+timezone: TIMEZONE \| TIMEZONE_OFFSET
	modifier: MODIFIER
	position: POSITION
	diff --git a/timefhuman/lalr.py b/timefhuman/lalr.py
	index 4b56fc3..eff8795 100644
	--- a/timefhuman/lalr.py
	+++ b/timefhuman/lalr.py
	@@ -1,9 +1,10 @@
	-from datetime import timedelta
	+from datetime import timedelta, timezone as dt_timezone
	from pathlib import Path
	import re

	from dataclasses import replace
	from lark import Lark, Transformer, UnexpectedInput
	+from lark.exceptions import VisitError
	import pytz
	from dateutil.relativedelta import relativedelta, weekdays

	@@ -14,6 +15,9 @@ from timefhuman.semantics import (
	DATE_TIME_NAME_TO_TEMPLATE,
	MODIFIER_TO_OFFSET,
	NUMBER_WORDS,
	+ build_date,
	+ build_numeric_date,
	+ build_time,
	POSITION_TO_DELTA,
	TIME_NAME_TO_TEMPLATE,
	UNIT_ALIASES,
	@@ -85,7 +89,12 @@ def parse_lalr_renderers(string: str, config: tfhConfig, start_pos: int = 0):
	except UnexpectedInput:
	return None

	- renderers = tfhTransformer(config=config).transform(tree)
	+ try:
	+ renderers = tfhTransformer(config=config).transform(tree)
	+ except VisitError as exc:
	+ if isinstance(exc.orig_exc, ValueError):
	+ return None
	+ raise
	for renderer in renderers:
	renderer.matched_text_pos = (span_start, span_end)
	return renderers
	@@ -186,16 +195,10 @@ class tfhTransformer(Transformer):
	sep = '/' if '/' in value else '-'
	parts = [int(part) for part in value.split(sep)]
	first, second = parts[0], parts[1]
	-
	- if len(parts) == 2:
	- if second > 31:
	- return {'date': tfhDate(month=first, year=normalize_year(second))}
	- return {'date': tfhDate(month=first, day=second)}
	-
	- third = parts[2]
	- if first >= 1000:
	- return {'date': tfhDate(year=first, month=second, day=third)}
	- return {'date': tfhDate(month=first, day=second, year=normalize_year(third))}
	+ result = build_numeric_date(first, second, parts[2] if len(parts) == 3 else None)
	+ if result is None:
	+ raise ValueError(f"Invalid numeric date: {value}")
	+ return {'date': result}

	def day(self, children):
	return {'day': int(children[0].value)}
	@@ -219,16 +222,28 @@ class tfhTransformer(Transformer):
	year = match.group('year')

	if year is not None:
	- return {'date': tfhDate(month=month, day=first, year=normalize_year(int(year)))}
	+ result = build_date(month=month, day=first, year=normalize_year(int(year)))
	+ if result is None:
	+ raise ValueError(f"Invalid monthname date: {children[0].value}")
	+ return {'date': result}
	if match.group('suffix') or first <= 31:
	- return {'date': tfhDate(month=month, day=first)}
	- return {'date': tfhDate(month=month, year=normalize_year(first))}
	+ result = build_date(month=month, day=first)
	+ if result is None:
	+ raise ValueError(f"Invalid monthname date: {children[0].value}")
	+ return {'date': result}
	+ result = build_date(month=month, year=normalize_year(first))
	+ if result is None:
	+ raise ValueError(f"Invalid monthname date: {children[0].value}")
	+ return {'date': result}

	def day_ordinal(self, children):
	match = DAY_ORDINAL_PATTERN.fullmatch(children[0].value)
	if match is None:
	raise NotImplementedError(f"Unknown day ordinal: {children[0].value}")
	- return {'date': tfhDate(day=int(match.group('day')))}
	+ result = build_date(day=int(match.group('day')))
	+ if result is None:
	+ raise ValueError(f"Invalid day ordinal: {children[0].value}")
	+ return {'date': result}

	def modified_month(self, children):
	offset = sum(child['offset'] for child in children[:-1])
	@@ -274,13 +289,16 @@ class tfhTransformer(Transformer):
	if 'time' in data:
	return {'time': data['time']}

	- return {'time': tfhTime(
	+ result = build_time(
	hour=int(data.get("hour", 0)),
	minute=int(data.get("minute", 0)),
	second=int(data.get("second", 0)),
	millisecond=int(data.get("millisecond", 0)),
	- meridiem=data.get("meridiem", None)
	- )}
	+ meridiem=data.get("meridiem", None),
	+ )
	+ if result is None:
	+ raise ValueError(f"Invalid time: {data}")
	+ return {'time': result}

	def meridiem(self, children):
	meridiem = children[0].value.lower()
	@@ -292,6 +310,14 @@ class tfhTransformer(Transformer):

	def timezone(self, children):
	timezone = children[0].value.lower()
	+ if timezone.startswith(("+", "-")):
	+ sign = -1 if timezone[0] == "-" else 1
	+ body = timezone[1:].replace(":", "")
	+ hours = int(body[:2])
	+ minutes = int(body[2:])
	+ if hours > 23 or minutes > 59:
	+ raise ValueError(f"Invalid timezone offset: {timezone}")
	+ return {'timezone': dt_timezone(sign * timedelta(hours=hours, minutes=minutes))}
	return {'timezone': pytz.timezone(timezone_mapping[timezone])}

	def timename(self, children):
	@@ -301,7 +327,10 @@ class tfhTransformer(Transformer):
	return {'time': clone_time(TIME_NAME_TO_TEMPLATE[timename])}

	def houronly(self, children):
	- return {'time': tfhTime(hour=int(children[0].value))}
	+ result = build_time(hour=int(children[0].value))
	+ if result is None:
	+ raise ValueError(f"Invalid hour: {children[0].value}")
	+ return {'time': result}

	def datetimename(self, children):
	datetimename = children[0].value.lower()
	@@ -326,7 +355,11 @@ class tfhTransformer(Transformer):
	if len(time_parts) > 2:
	second = time_parts[2]

	+ date = build_date(year=year, month=month, day=day)
	+ time = build_time(hour=hour, minute=minute, second=second, millisecond=millisecond)
	+ if date is None or time is None:
	+ raise ValueError(f"Invalid ISO datetime: {value}")
	return {'datetime': tfhDatetime(
	- date=tfhDate(year=year, month=month, day=day),
	- time=tfhTime(hour=hour, minute=minute, second=second, millisecond=millisecond),
	+ date=date,
	+ time=time,
	)}
	diff --git a/timefhuman/main.py b/timefhuman/main.py
	index 784f72e..d285e2f 100644
	--- a/timefhuman/main.py
	+++ b/timefhuman/main.py
	@@ -57,10 +57,25 @@ def _parse_renderers(string: str, config: tfhConfig):


	def _matched_results(string: str, renderers, config: tfhConfig):
	- positions = [(renderer.matched_text_pos[0], renderer.matched_text_pos[1]) for renderer in renderers]
	- matched_texts = [string[start:end] for start, end in positions]
	- datetimes = [renderer.to_object(config) for renderer in renderers]
	- return list(zip(matched_texts, positions, datetimes))
	+ results = []
	+ for renderer in renderers:
	+ try:
	+ value = renderer.to_object(config)
	+ except ValueError:
	+ continue
	+ start, end = renderer.matched_text_pos
	+ results.append((string[start:end], (start, end), value))
	+ return results
	+
	+
	+def _valid_objects(renderers, config: tfhConfig):
	+ results = []
	+ for renderer in renderers:
	+ try:
	+ results.append(renderer.to_object(config))
	+ except ValueError:
	+ continue
	+ return results


	def build_raw_tree(string: str, config: tfhConfig):
	@@ -97,4 +112,4 @@ def timefhuman(string, config: tfhConfig = DEFAULT_CONFIG, raw: bool=False, now:

	if config.return_matched_text:
	return _matched_results(string, renderers, config)
	- return [renderer.to_object(config) for renderer in renderers]
	+ return _valid_objects(renderers, config)
	diff --git a/timefhuman/semantics.py b/timefhuman/semantics.py
	index bfc6317..cbb5d9e 100644
	--- a/timefhuman/semantics.py
	+++ b/timefhuman/semantics.py
	@@ -1,8 +1,8 @@
	-from datetime import timedelta
	+from datetime import date, timedelta

	from dateutil.relativedelta import relativedelta

	-from timefhuman.renderers import tfhDatetime, tfhTime
	+from timefhuman.renderers import tfhDate, tfhDatetime, tfhTime
	from timefhuman.utils import get_month_mapping


	@@ -103,6 +103,69 @@ def normalize_year(value: int):
	return value


	+def build_date(year: int \| None = None, month: int \| None = None, day: int \| None = None):
	+ if year is not None and not 1 <= year <= 9999:
	+ return None
	+ if month is not None and not 1 <= month <= 12:
	+ return None
	+ if day is not None and not 1 <= day <= 31:
	+ return None
	+ if month is not None and day is not None:
	+ validation_year = year if year is not None else 2000
	+ try:
	+ date(validation_year, month, day)
	+ except ValueError:
	+ return None
	+ return tfhDate(year=year, month=month, day=day)
	+
	+
	+def build_numeric_date(first: int, second: int, third: int \| None = None):
	+ if third is None:
	+ if first >= 1000:
	+ return build_date(year=first, month=second)
	+ if second > 31:
	+ return build_date(month=first, year=normalize_year(second))
	+ if 1 <= first <= 12:
	+ return build_date(month=first, day=second)
	+ if 1 <= second <= 12:
	+ return build_date(month=second, day=first)
	+ return None
	+
	+ if first >= 1000:
	+ return build_date(year=first, month=second, day=third)
	+
	+ year = normalize_year(third)
	+ if 1 <= first <= 12:
	+ return build_date(month=first, day=second, year=year)
	+ if 1 <= second <= 12:
	+ return build_date(year=year, month=second, day=first)
	+ return None
	+
	+
	+def build_time(
	+ hour: int,
	+ minute: int = 0,
	+ second: int = 0,
	+ millisecond: int = 0,
	+ meridiem=None,
	+):
	+ if not 0 <= minute <= 59 or not 0 <= second <= 59 or not 0 <= millisecond <= 999999:
	+ return None
	+ if meridiem is None:
	+ if not 0 <= hour <= 23:
	+ return None
	+ elif not 1 <= hour <= 12:
	+ return None
	+
	+ return tfhTime(
	+ hour=hour,
	+ minute=minute,
	+ second=second,
	+ millisecond=millisecond,
	+ meridiem=meridiem,
	+ )
	+
	+
	def clone_time(template: tfhTime):
	return tfhTime(
	hour=template.hour,