Upload convert_tot.py with huggingface_hub

convert_tot.py

@@ -0,0 +1,72 @@
+import ast, hashlib, json, random, re
+from collections import Counter
+from itertools import chain
+from datasets import load_dataset
+
+random.seed(42)
+MAX_ROWS = 42000
+rows, per_qt, fps = [], Counter(), set()
+
+ds = chain(load_dataset("baharef/ToT", "tot_arithmetic", split="test", streaming=True),
+           load_dataset("baharef/ToT", "tot_semantic", split="test", streaming=True),
+           load_dataset("baharef/ToT", "tot_semantic_large", split="test", streaming=True))
+
+NUM = re.compile(r"-?\d+")
+for r in ds:
+    qt = r.get("question_type") or "misc"
+    q = str(r.get("question", "")).strip()
+    p = str(r.get("prompt", "")).strip()
+    lab = r.get("label")
+    if isinstance(lab, str):
+        try:
+            lab = ast.literal_eval(lab)
+        except Exception:
+            pass
+    ans = lab.get("answer") if isinstance(lab, dict) else (lab if isinstance(lab, str) else None)
+    if not q or not ans or len(p) > 1400:
+        continue
+    ans = str(ans).strip()
+    m = NUM.search(ans.replace(",", ""))
+    if not m:
+        continue
+    val = int(m.group(0))
+    # wrong candidates: same answer string, first number perturbed
+    def swap(a, v):
+        return NUM.sub(str(v), a, count=1)
+    cands = {val}
+    for d in (1, -1, 2, -2, 10, -10, 4, -4, 8, -8, 100, -100, 5, -5):
+        cands.add(val + d)
+        if len(cands) >= 4:
+            break
+    opts = [ans] + [swap(ans.replace(",", ""), x) for x in random.sample(sorted(cands - {val}), 3)]
+    random.shuffle(opts)
+    fp = hashlib.sha256(f"{p}|{q}|{ans}".encode()).hexdigest()
+    if fp in fps:
+        continue
+    fps.add(fp)
+    qq = re.sub(r"format your answer.*$", "", q, flags=re.I).strip()
+    st = {"context": p} if p else {"context": qq}
+    tgt = [0.0] * len(opts)
+    ti = opts.index(ans)
+    tgt[ti] = 1.0
+    gid = f"tot:{qt}:{fp[:12]}"
+    rows.append({"state": st, "id": f"tot:{qt}:{fp[:16]}", "kind": "choice",
+                 "options": opts, "target": tgt,
+                 "question": qq + " Choose the correct answer.",
+                 "source": f"tot:{qt}", "group_id": gid})
+    per_qt[qt] += 1
+    if len(rows) >= MAX_ROWS:
+        break
+
+random.shuffle(rows)
+dev = [r for r in rows if hashlib.md5(r["group_id"].encode()).hexdigest()[0] == "0"]
+train = [r for r in rows if hashlib.md5(r["group_id"].encode()).hexdigest()[0] != "0"]
+tg = {r["group_id"] for r in train}
+dev = [r for r in dev if r["group_id"] not in tg]
+with open("/root/alea/data/tot_train.jsonl", "w") as f:
+    for r in train[:36000]:
+        f.write(json.dumps(r) + "\n")
+with open("/root/alea/data/tot_dev.jsonl", "w") as f:
+    for r in dev[:3500]:
+        f.write(json.dumps(r) + "\n")
+print({"train": min(36000, len(train)), "dev": min(3500, len(dev)), "qtypes": dict(per_qt)})