Add evaluation-only diagnostic for recovery candidate

training/check_recovery.sh

@@ -0,0 +1,69 @@
+#!/usr/bin/env bash
+set -euo pipefail
+: "${HF_TOKEN:?기존 Colab의 HF_TOKEN을 설정하세요.}"
+export HF_HUB_DISABLE_XET=1 HF_HUB_DISABLE_TELEMETRY=1 TOKENIZERS_PARALLELISM=false DCT_COMPILE=0
+export PICK_TRAIN_ROOT="${PICK_TRAIN_ROOT:-/content/pick-recovery}"
+PICK_PYTHON="$PICK_TRAIN_ROOT/.venv/bin/python3"
+if [[ ! -x "$PICK_PYTHON" || ! -d "$PICK_TRAIN_ROOT/base-model" ]]; then
+  echo '학습을 완료한 기존 Colab 런타임에서 실행하세요. 평가 환경을 찾지 못했습니다.' >&2
+  exit 1
+fi
+"$PICK_PYTHON" - <<'PY'
+import hashlib, json, os, shutil
+from pathlib import Path
+from huggingface_hub import HfApi, hf_hub_download
+root=Path(os.environ['PICK_TRAIN_ROOT'])
+repo='lCREEl/pick-v6-recovery'
+revision='7f06104ea66066fae5bedc90e914bdd94e50e0a6'
+assert HfApi().repo_info(repo).private, 'Diagnostic outputs must stay private'
+expected='a723199fce99ca00495f87ad1304e822225af0adc7e9a16dd17916abe0c2a995'
+checkpoint=root/'runs/frozen-head/latest.pt'
+if not checkpoint.exists():
+    source=hf_hub_download(repo,'runs/frozen-head/latest.pt',revision=revision)
+    checkpoint.parent.mkdir(parents=True,exist_ok=True)
+    shutil.copyfile(source,checkpoint)
+with checkpoint.open('rb') as handle:
+    assert hashlib.file_digest(handle,'sha256').hexdigest()==expected, 'Unexpected checkpoint'
+manifest=json.loads((root/'base-model-manifest.json').read_text())
+for name, sha in manifest.items():
+    with (root/'base-model'/name).open('rb') as handle:
+        assert hashlib.file_digest(handle,'sha256').hexdigest()==sha, name
+print('V5와 600단계 후보 해시 검증 완료. 별도 312문항 진단 평가를 시작합니다.',flush=True)
+PY
+"$PICK_PYTHON" "$PICK_TRAIN_ROOT/scripts/run_pilot.py" \
+  --export "$PICK_TRAIN_ROOT/base-model" --device cuda --eval-only --confirm \
+  --name frozen-head-diagnostic-v5
+"$PICK_PYTHON" "$PICK_TRAIN_ROOT/scripts/run_pilot.py" \
+  --export "$PICK_TRAIN_ROOT/base-model" --device cuda --eval-only --confirm \
+  --checkpoint "$PICK_TRAIN_ROOT/runs/frozen-head/latest.pt" \
+  --name frozen-head-diagnostic-candidate
+"$PICK_PYTHON" - <<'PY'
+import json, os, tempfile
+from pathlib import Path
+from huggingface_hub import HfApi
+root=Path(os.environ['PICK_TRAIN_ROOT'])/'runs'
+names=['frozen-head-diagnostic-v5','frozen-head-diagnostic-candidate']
+base, candidate=[json.loads((root/name/'confirm.json').read_text()) for name in names]
+assert [r['id'] for r in base['records']]==[r['id'] for r in candidate['records']]
+deltas={task:100*(candidate['per_task'][task]['score']-values['score'])
+        for task,values in base['per_task'].items()}
+report=dict(status='diagnostic_only',original_development_decision='discard',
+    baseline_proxy=base['proxy_score'],candidate_proxy=candidate['proxy_score'],
+    delta=candidate['proxy_score']-base['proxy_score'],per_task_delta_points=deltas,
+    worst_anchor_delta_points=min(deltas[t] for t in ['banking','clinc','home']),
+    public_index=None,
+    note='Separate diagnostic after a failed development gate. Does not promote the candidate or change submitted V5.')
+out=root/'frozen-head-diagnostic';out.mkdir(exist_ok=True)
+(out/'summary.json').write_text(json.dumps(report,ensure_ascii=False,indent=2)+'\n')
+# Upload only diagnostic JSONs, in one commit. Preserve original decision and checkpoint.
+with tempfile.TemporaryDirectory() as staging:
+    for name in names+['frozen-head-diagnostic']:
+        target=Path(staging)/name;target.mkdir()
+        for path in (root/name).glob('*.json'):
+            (target/path.name).write_bytes(path.read_bytes())
+    api=HfApi();assert api.repo_info('lCREEl/pick-v6-recovery').private
+    api.upload_folder(repo_id='lCREEl/pick-v6-recovery',folder_path=staging,
+        path_in_repo='runs',commit_message='Record diagnostic confirmation; preserve development discard')
+print(json.dumps(report,ensure_ascii=False,indent=2),flush=True)
+print('진단 결과 HF 백업 완료. 추가 학습 없이 종료했습니다.',flush=True)
+PY