GuanyuH commited on
Commit
b56998f
路
verified 路
1 Parent(s): e672914

Load tasks from the published dataset (v1.0)

Browse files
Files changed (1) hide show
  1. app.py +97 -5
app.py CHANGED
@@ -49,14 +49,106 @@ GA_HEAD = """<script async src="https://www.googletagmanager.com/gtag/js?id=G-W0
49
  # ]
50
  # }, ...
51
  # }
52
- # Swap this for a Hugging Face Dataset load when you outgrow a local file.
53
- with open("tasks.json") as f:
54
- TASKS = json.load(f)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55
 
56
  # Friendly dropdown: (label, value) so the menu reads as titles, not opaque IDs.
 
57
  TASK_CHOICES = [
58
- (f"{tid} 路 {t.get('doc_title', '')}", tid)
59
- for tid, t in TASKS.items()
 
 
 
 
60
  ]
61
 
62
 
 
49
  # ]
50
  # }, ...
51
  # }
52
+ # The task bank is the published dataset:
53
+ # https://huggingface.co/datasets/CentificAIResearch/Healthcare.pdf
54
+ # It is downloaded once and cached. The bundled tasks.json is kept only as a
55
+ # fallback so a Hub outage degrades the Space to the local demo tasks instead of
56
+ # taking it down. Set HEALTHCARE_PDF_LOCAL=1 to force the local file.
57
+ DATASET_REPO = os.environ.get("HEALTHCARE_PDF_DATASET", "CentificAIResearch/Healthcare.pdf")
58
+ DATASET_REVISION = os.environ.get("HEALTHCARE_PDF_REVISION", "v1.0")
59
+
60
+
61
+ def _tasks_from_hub():
62
+ """Map the published dataset onto the task-bank shape documented above.
63
+
64
+ Anchors live in their own column keyed by id; each sub-question references
65
+ them by `anchor_ids`, so they are resolved here into the display strings the
66
+ UI expects ("a1: <quote>").
67
+ """
68
+ import pandas as pd
69
+ from huggingface_hub import snapshot_download
70
+
71
+ root = snapshot_download(DATASET_REPO, repo_type="dataset", revision=DATASET_REVISION)
72
+ cols = ["task_id", "persona", "occupation", "aspect", "doc_id", "document_name",
73
+ "document_file", "case_setup", "task_prompt", "sub_questions", "anchors",
74
+ "difficulty_band", "n_models_solved"]
75
+ # `document` (the embedded PDF bytes) is deliberately not read -- the same
76
+ # files are on disk under pdfs/ and fitz needs a path, not bytes.
77
+ df = pd.read_parquet(os.path.join(root, "data", "test.parquet"), columns=cols)
78
+
79
+ tasks = {}
80
+ for r in df.itertuples():
81
+ by_id = {a["id"]: a for a in r.anchors}
82
+ tasks[r.task_id] = {
83
+ "occupation": r.occupation,
84
+ "persona": r.persona,
85
+ "aspect": r.aspect,
86
+ "doc_pid": r.doc_id,
87
+ "doc_title": r.document_name,
88
+ "pdf_path": os.path.join(root, "pdfs", r.document_file),
89
+ "prompt": r.task_prompt or r.case_setup,
90
+ "case_setup": r.case_setup,
91
+ "difficulty_band": r.difficulty_band,
92
+ "n_models_solved": int(r.n_models_solved),
93
+ "sub_questions": [
94
+ {"q_no": int(s["q_no"]),
95
+ "question": s["question"],
96
+ "reference_answer": s["reference_answer"],
97
+ "task_types": list(s["task_types"]),
98
+ "anchors": [f"{aid}: {by_id[aid]['locator_quote']}"
99
+ for aid in s["anchor_ids"] if aid in by_id]}
100
+ for s in r.sub_questions
101
+ ],
102
+ }
103
+ return tasks
104
+
105
+
106
+ def _load_tasks():
107
+ if os.environ.get("HEALTHCARE_PDF_LOCAL") == "1":
108
+ with open("tasks.json") as f:
109
+ return json.load(f), "local (forced)"
110
+ try:
111
+ return _tasks_from_hub(), f"{DATASET_REPO}@{DATASET_REVISION}"
112
+ except Exception as exc: # noqa: BLE001 -- never take the Space down over this
113
+ print(f"[tasks] Hub load failed ({type(exc).__name__}: {exc}); "
114
+ f"falling back to bundled tasks.json")
115
+ with open("tasks.json") as f:
116
+ return json.load(f), "local (fallback)"
117
+
118
+
119
+ TASKS, TASK_SOURCE = _load_tasks()
120
+ print(f"[tasks] loaded {len(TASKS)} tasks from {TASK_SOURCE}")
121
+
122
+ _OCC_SHORT = {
123
+ "emergency_medicine_physicians": "Emergency Medicine",
124
+ "family_medicine_physicians": "Family Medicine",
125
+ "general_internal_medicine_physicians": "General Internal Med",
126
+ "nurse_practitioners": "Nurse Practitioners",
127
+ "physician_assistants": "Physician Assistants",
128
+ "pharmacists": "Pharmacists",
129
+ "pharmacy_technicians": "Pharmacy Technicians",
130
+ }
131
+ _BAND_ORDER = {"unsolved": 0, "hard": 1, "mid": 2, "easy": 3, "ceiling": 4}
132
+
133
+
134
+ def _task_label(tid, t):
135
+ """Dropdown label: difficulty, occupation, then the document it is grounded in."""
136
+ occ = _OCC_SHORT.get(t.get("occupation", ""), t.get("occupation", ""))
137
+ band = t.get("difficulty_band")
138
+ head = f"[{band}] " if band else ""
139
+ title = (t.get("doc_title") or "")[:60]
140
+ return f"{head}{occ} 路 {title}" if occ else f"{tid} 路 {title}"
141
+
142
 
143
  # Friendly dropdown: (label, value) so the menu reads as titles, not opaque IDs.
144
+ # Hardest first, so the interesting cases are at the top of the list.
145
  TASK_CHOICES = [
146
+ (_task_label(tid, t), tid)
147
+ for tid, t in sorted(
148
+ TASKS.items(),
149
+ key=lambda kv: (_BAND_ORDER.get(kv[1].get("difficulty_band"), 9),
150
+ kv[1].get("occupation", ""), kv[0]),
151
+ )
152
  ]
153
 
154