structverify 0.3.1__tar.gz → 0.3.3__tar.gz

{structverify-0.3.1/structverify.egg-info → structverify-0.3.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: structverify
-Version: 0.3.1
+Version: 0.3.3
 Summary: Compliance & fact verification for documents — check your files against your own rulebook (PDF) or a data source, and read a plain True/False.
 Author-email: "김예슬 (Yeseul Kim)" <yesul0718@gmail.com>
 License: MIT License

{structverify-0.3.1 → structverify-0.3.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "structverify"
-version = "0.3.1"
+version = "0.3.3"
 description = "Compliance & fact verification for documents — check your files against your own rulebook (PDF) or a data source, and read a plain True/False."
 readme = "README.md"
 license = { file = "LICENSE" }

{structverify-0.3.1 → structverify-0.3.3}/structverify/__init__.py

@@ -54,7 +54,7 @@ _TRAINING_SYMBOLS = frozenset({
 })
 
 __all__ = sorted(_API_SYMBOLS | _PIPELINE_SYMBOLS | _PROGRESS_SYMBOLS | _TRAINING_SYMBOLS)
-__version__ = "0.3.1"
+__version__ = "0.3.3"
 
 
 def __getattr__(name: str):

{structverify-0.3.1 → structverify-0.3.3}/structverify/training/loop.py

@@ -96,7 +96,11 @@ class LearningLoop:
             print("   " + info["command"])
             return info
         try:
-            __import__(_probe)
+            # 백엔드 존재 확인 — torch 등 무거운 import가 뿜는 잡음(torchao·Flax)은 숨긴다.
+            import contextlib
+            with open(os.devnull, "w") as _dn, \
+                 contextlib.redirect_stdout(_dn), contextlib.redirect_stderr(_dn):
+                __import__(_probe)
         except Exception:  # noqa: BLE001
             raise RuntimeError(f"run=True에는 학습 백엔드가 필요합니다 → {need}. "
                                "또는 run=False로 명령만 받아 별도 실행하세요.")

{structverify-0.3.1 → structverify-0.3.3}/structverify/training/monitor.py

@@ -10,11 +10,34 @@
 """
 from __future__ import annotations
 
+import os
+import sys
 import time
 from dataclasses import dataclass, field
 
 _SPARK = "▁▂▃▄▅▆▇█"
 
+# ── 터미널 색 (v0.3.2 팔레트) ────────────────────────────────────────────────
+_ANSI = {"cyan": "\x1b[38;5;117m", "spark": "\x1b[38;5;111m", "dim": "\x1b[2m",
+         "bold": "\x1b[1m", "green": "\x1b[38;5;114m", "amber": "\x1b[38;5;179m",
+         "red": "\x1b[38;5;210m", "reset": "\x1b[0m"}
+
+
+def _color_enabled() -> bool:
+    """NO_COLOR면 끔, SV_COLOR=1이면 강제 켬, 그 외엔 tty 감지."""
+    if os.environ.get("NO_COLOR"):
+        return False
+    if os.environ.get("SV_COLOR"):
+        return True
+    try:
+        return sys.stderr.isatty()
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _c(name: str, s: str, on: bool) -> str:
+    return f"{_ANSI[name]}{s}{_ANSI['reset']}" if on else s
+
 
 def sparkline(values: list[float], width: int = 8) -> str:
     """최근 값들을 ▁▂▃ 스파크라인으로. 값이 없으면 빈 문자열."""
@@ -118,19 +141,29 @@ class LossMonitor:
 
     # ── 터미널 UI ───────────────────────────────────────────────────────────
     def render(self, max_steps: int, started_at: float, lr: float | None = None,
-               width: int = 12) -> str:
-        """현재 상태를 터미널 한 줄로."""
+               width: int = 12, color: bool | None = None) -> str:
+        """현재 상태를 터미널 한 줄로 (tty면 색 포함 — NO_COLOR/SV_COLOR로 제어)."""
+        on = _color_enabled() if color is None else color
         frac = self.step / max_steps if max_steps else 0.0
         filled = int(frac * width)
         bar = "▓" * filled + "░" * (width - filled)
         loss = self.losses[-1] if self.losses else float("nan")
         spark = sparkline(self.means)
         elapsed = time.time() - started_at
-        eta = _fmt_eta(elapsed / frac - elapsed) if frac > 0.01 else "--:--"
-        lr_s = f" · lr {lr:.1e}" if lr else ""
-        spike_s = f" · ⚡{len(self.spikes)}" if self.spikes else ""
-        return (f"🧠 {self.step}/{max_steps} {bar} {frac*100:3.0f}% · "
-                f"loss {loss:.3f} {spark}{lr_s} · ETA {eta}{spike_s} · 🩺 {self.status}")
+        eta = _fmt_eta(elapsed / frac - elapsed) if frac > 0.005 else "--:--"
+        st_color = ("red" if "발산" in self.status
+                    else "amber" if self.status == "워밍업" else "green")
+        parts = [
+            "🧠 " + _c("cyan", f"{self.step}/{max_steps} {bar} {frac*100:3.0f}%", on),
+            _c("bold", f"loss {loss:.3f}", on) + " " + _c("spark", spark, on),
+        ]
+        if lr:
+            parts.append(_c("dim", f"lr {lr:.1e}", on))
+        parts.append(_c("dim", f"ETA {eta}", on))
+        if self.spikes:
+            parts.append(_c("amber", f"⚡{len(self.spikes)}", on))
+        parts.append("🩺 " + _c(st_color, self.status, on))
+        return " · ".join(parts)
 
     def summary(self) -> str:
         """학습 종료 시 요약 (레시피가 마지막에 출력)."""

{structverify-0.3.1 → structverify-0.3.3}/structverify/training/recipe/sample.py

@@ -9,10 +9,27 @@ GPU 머신(어댑터 학습한 곳)에서 실행:
         --task schema --claim "당사 물류 파트너는 2,300개사에 달했습니다."
 """
 import argparse
+import contextlib
+import logging
 import os
+import sys
+import warnings
 
 os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "3")
 os.environ.setdefault("TRANSFORMERS_VERBOSITY", "error")
+sys.unraisablehook = lambda *_: None
+warnings.filterwarnings("ignore")
+logging.getLogger("torchao").setLevel(logging.CRITICAL)
+
+
+@contextlib.contextmanager
+def _quiet():
+    if os.environ.get("SV_VERBOSE"):
+        yield
+        return
+    with open(os.devnull, "w") as dn, \
+         contextlib.redirect_stdout(dn), contextlib.redirect_stderr(dn):
+        yield
 
 # 내장 테스트 — 시드/증강 데이터에 없는 새 표현으로 일반화를 확인.
 _DEFAULT_TESTS = [
@@ -37,7 +54,8 @@ def main():
     ap.add_argument("--max-tokens", type=int, default=120)
     args = ap.parse_args()
 
-    from unsloth import FastLanguageModel
+    with _quiet():
+        from unsloth import FastLanguageModel
 
     from structverify.training.tasks import build_example
 
@@ -49,10 +67,11 @@ def main():
     else:
         tests = _DEFAULT_TESTS
 
-    print(f"어댑터 로드: {args.adapter}")
-    model, tok = FastLanguageModel.from_pretrained(
-        args.adapter, max_seq_length=2048, load_in_4bit=True)
-    FastLanguageModel.for_inference(model)
+    print(f"⚙ 어댑터 로드 — {args.adapter}")
+    with _quiet():
+        model, tok = FastLanguageModel.from_pretrained(
+            args.adapter, max_seq_length=2048, load_in_4bit=True)
+        FastLanguageModel.for_inference(model)
 
     for t in tests:
         msgs = build_example(**t)["messages"][:2]     # system + user만

{structverify-0.3.1 → structverify-0.3.3}/structverify/training/recipe/train_qlora.py

@@ -26,6 +26,29 @@ os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "3")
 os.environ.setdefault("TRANSFORMERS_VERBOSITY", "error")
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 
+# "Exception ignored in: <_io.BytesIO ...> BufferError" 도배 차단.
+#   Python 3.13 + datasets 멀티프로세스 토큰화의 무해한 정리 단계 잡음인데,
+#   워커가 fork로 뜨면서 이 훅을 물려받아 자식 프로세스 출력까지 조용해진다.
+sys.unraisablehook = lambda *_: None
+
+import contextlib
+import logging
+import warnings
+
+warnings.filterwarnings("ignore")                       # Flax deprecated 등
+logging.getLogger("torchao").setLevel(logging.CRITICAL)  # torchao .so 로드 WARNING
+
+
+@contextlib.contextmanager
+def _quiet():
+    """서드파티 import·로드가 뿜는 무관한 출력 숨김. SV_VERBOSE=1이면 전부 표시."""
+    if os.environ.get("SV_VERBOSE"):
+        yield
+        return
+    with open(os.devnull, "w") as dn, \
+         contextlib.redirect_stdout(dn), contextlib.redirect_stderr(dn):
+        yield
+
 
 def load_chat_dataset(path):
     from datasets import Dataset
@@ -61,26 +84,49 @@ def main():
                     help="중간 체크포인트(checkpoint-*) 보존 — 기본은 성공 시 삭제 (어댑터만 수십 MB)")
     args = ap.parse_args()
 
-    # unsloth: 3060/T4에서 QLoRA 2배 빠름 + 메모리 절약
-    from unsloth import FastLanguageModel
-    from trl import SFTTrainer, SFTConfig
-
-    model, tokenizer = FastLanguageModel.from_pretrained(
-        model_name=args.model,
-        max_seq_length=args.seq_len,
-        load_in_4bit=True,          # QLoRA — 12GB VRAM에 7B 적재
-    )
-    model = FastLanguageModel.get_peft_model(
-        model,
-        r=args.rank,
-        target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
-                        "gate_proj", "up_proj", "down_proj"],
-        lora_alpha=args.rank * 2,
-        lora_dropout=0.0,
-        bias="none",
-        use_gradient_checkpointing="unsloth",
-    )
-
+    # unsloth: 3060/T4에서 QLoRA 2배 빠름 + 메모리 절약.
+    # import·로드 단계의 배너·경고는 숨기고(_quiet) 우리 요약 한 줄로 대체.
+    with _quiet():
+        from unsloth import FastLanguageModel
+        from trl import SFTTrainer, SFTConfig
+
+        # tqdm 진행바(체크포인트 로드·Map·Tokenizing) 대신 우리 phase 라인으로.
+        try:
+            import datasets as _ds
+            import transformers as _tf
+            _ds.disable_progress_bars()
+            _tf.utils.logging.disable_progress_bar()
+        except Exception:  # noqa: BLE001
+            pass
+
+    try:
+        import torch
+        _p = torch.cuda.get_device_properties(0)
+        _prec = "bf16" if torch.cuda.is_bf16_supported() else "fp16"
+        print(f"⚙ 학습 엔진 준비 — unsloth · {_p.name} ({_p.total_memory/2**30:.0f}GB) · {_prec}",
+              file=sys.stderr)
+    except Exception:  # noqa: BLE001
+        pass
+
+    print(f"⚙ 베이스 모델 로드 — {args.model} (4bit)", file=sys.stderr)
+    with _quiet():
+        model, tokenizer = FastLanguageModel.from_pretrained(
+            model_name=args.model,
+            max_seq_length=args.seq_len,
+            load_in_4bit=True,          # QLoRA — 12GB VRAM에 7B 적재
+        )
+        model = FastLanguageModel.get_peft_model(
+            model,
+            r=args.rank,
+            target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
+                            "gate_proj", "up_proj", "down_proj"],
+            lora_alpha=args.rank * 2,
+            lora_dropout=0.0,
+            bias="none",
+            use_gradient_checkpointing="unsloth",
+        )
+
+    print("⚙ 데이터 준비·토큰화 중 …", file=sys.stderr)
     ds = load_chat_dataset(args.data)
 
     def fmt(ex):
@@ -112,7 +158,7 @@ def main():
     from transformers import TrainerCallback
     from transformers.trainer_callback import PrinterCallback
 
-    from structverify.training.monitor import LossMonitor
+    from structverify.training.monitor import LossMonitor, _c, _color_enabled
 
     class SVLiveCallback(TrainerCallback):
         def __init__(self, early_stop: bool, patience: int):
@@ -128,18 +174,20 @@ def main():
             print("\r\x1b[2K" + line, end="", flush=True, file=sys.stderr)
             if self.mon.should_stop and self.early_stop:
                 print("", file=sys.stderr)
-                print(f"✂ {self.mon.stop_reason}", file=sys.stderr)
+                print(_c("amber", f"✂ {self.mon.stop_reason}", _color_enabled()), file=sys.stderr)
                 control.should_training_stop = True
             return control
 
         def on_train_end(self, targs, state, control, **kw):
             print("", file=sys.stderr)
-            print(self.mon.summary(), file=sys.stderr)
+            print(_c("green", self.mon.summary(), _color_enabled()), file=sys.stderr)
             return control
 
     trainer.remove_callback(PrinterCallback)     # {'loss': ...} dict 도배 제거
     trainer.add_callback(SVLiveCallback(args.early_stop, args.patience))
 
+    es = "ON" if args.early_stop else "OFF (관찰만)"
+    print(f"🩺 실시간 감독 시작 — patience {args.patience} · 조기 종료 {es}", file=sys.stderr)
     trainer.train()
 
     # 어댑터 + trainer_state.json 저장 (TrainDoctor가 읽음)

{structverify-0.3.1 → structverify-0.3.3/structverify.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: structverify
-Version: 0.3.1
+Version: 0.3.3
 Summary: Compliance & fact verification for documents — check your files against your own rulebook (PDF) or a data source, and read a plain True/False.
 Author-email: "김예슬 (Yeseul Kim)" <yesul0718@gmail.com>
 License: MIT License