Upload scripts/convert_reasoning.py with huggingface_hub

scripts/convert_reasoning.py

@@ -0,0 +1,123 @@
+"""Convert BIG-bench (tasksource mirror) + LogiQA + ReClor into Alea canonical rows.
+All are MCQ with exact labels -> choice rows. Group-disjoint dev via hash split."""
+import hashlib, json, random, re
+from collections import Counter
+from datasets import load_dataset
+
+random.seed(7)
+D = "/root/alea/data/"
+rows, per_src, fps = [], Counter(), set()
+
+BB_CONFIGS = ["logical_deduction", "tracking_shuffled_objects", "web_of_lies",
+              "boolean_expressions", "date_understanding", "temporal_sequences",
+              "multistep_arithmetic_two", "navigate", "hyperbaton", "dyck_languages",
+              "reasoning_about_colored_objects", "penguins_in_a_table",
+              "formal_fallacies", "minute_mysteries_qa", "snarks",
+              "analytic_entailment", "general_knowledge", "disambiguation_qa"]
+CAP = 2600
+
+def gid(*xs):
+    return hashlib.sha256("|".join(xs).encode()).hexdigest()[:16]
+
+for cfg in BB_CONFIGS:
+    n = 0
+    for split in ("train", "validation"):
+        try:
+            ds = load_dataset("tasksource/bigbench", cfg, split=split, streaming=True)
+        except Exception as e:
+            print(cfg, split, "ERR", str(e)[:60]); continue
+        for r in ds:
+            if n >= CAP:
+                break
+            opts = [str(o).strip() for o in (r.get("multiple_choice_targets") or [])]
+            sc = r.get("multiple_choice_scores") or []
+            if len(opts) < 2 or len(opts) != len(sc) or len(opts) > 14:
+                # non-MCQ: fall back to noul verification on the text target
+                t = r.get("targets") or []
+                if not t:
+                    continue
+                ans = str(t[0]).strip()
+                if len(ans) > 60 or not ans:
+                    continue
+                correct = random.random() < 0.5
+                shown = ans if correct else (ans + "x" if ans else "x")
+                if not correct and shown == ans:
+                    continue
+                state = {"task": cfg.replace("_", " "), "problem": str(r["inputs"]).strip()[:1800]}
+                qs = {"q": {"type": "noul",
+                            "instructions": f"Is the proposed answer correct? Proposed answer: {shown}",
+                            "criteria": {}}}
+                tg = {"q": float(correct)}
+                kind = "noul"
+            else:
+                ti = sc.index(max(sc))
+                state = {"task": cfg.replace("_", " "), "problem": str(r["inputs"]).strip()[:1800]}
+                crit = {chr(65 + i): o for i, o in enumerate(opts)}
+                qs = {"q": {"type": "choice",
+                            "instructions": "Choose the correct answer.",
+                            "criteria": crit}}
+                vec = [0.0] * len(opts); vec[ti] = 1.0
+                tg = {"q": vec}
+            g = f"bb:{cfg}:{r.get('idx', n)}"
+            fp = gid(state["problem"], json.dumps(qs, sort_keys=True))
+            if fp in fps:
+                continue
+            fps.add(fp)
+            rows.append({"state": state, "id": g, "questions": qs, "targets": tg,
+                         "source": f"bigbench:{cfg}", "group_id": g,
+                         "family": f"bigbench:{cfg}", "split0": split})
+            n += 1
+            per_src[cfg] += 1
+    print(cfg, per_src[cfg])
+
+# LogiQA
+try:
+    for r in load_dataset("lucasmccabe/logiqa", split="train", streaming=True):
+        opts = [str(o).strip() for o in r["options"]]
+        if len(opts) < 2 or len(opts) > 10:
+            continue
+        try:
+            ti = int(r["correct_option"])
+        except Exception:
+            continue
+        state = {"context": str(r["context"]).strip()[:1400],
+                 "question": str(r["query"]).strip()}
+        crit = {chr(65 + i): o for i, o in enumerate(opts)}
+        vec = [0.0] * len(opts); vec[ti] = 1.0
+        g = f"logiqa:{r.get('id', len(rows))}"
+        rows.append({"state": state, "id": g,
+                     "questions": {"q": {"type": "choice", "instructions": "Choose the logically correct answer.", "criteria": crit}},
+                     "targets": {"q": vec}, "source": "logiqa", "group_id": g,
+                     "family": "logiqa"})
+        per_src["logiqa"] += 1
+except Exception as e:
+    print("logiqa ERR", e)
+
+# ReClor
+try:
+    for r in load_dataset("metaeval/reclor", split="train", streaming=True):
+        opts = [str(o).strip() for o in r["answers"]]
+        lab = r.get("label")
+        if lab is None or lab < 0 or len(opts) != 4:
+            continue
+        state = {"context": str(r["context"]).strip()[:1400],
+                 "question": str(r["question"]).strip()}
+        crit = {chr(65 + i): o for i, o in enumerate(opts)}
+        vec = [0.0] * 4; vec[lab] = 1.0
+        g = f"reclor:{r.get('id_string', len(rows))}"
+        rows.append({"state": state, "id": g,
+                     "questions": {"q": {"type": "choice", "instructions": "Choose the logically correct answer.", "criteria": crit}},
+                     "targets": {"q": vec}, "source": "reclor", "group_id": g,
+                     "family": "reclor"})
+        per_src["reclor"] += 1
+except Exception as e:
+    print("reclor ERR", e)
+
+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]
+open(D + "reason_train.jsonl", "w").write("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in train))
+open(D + "reason_dev.jsonl", "w").write("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in dev[:4000]))
+print("train", len(train), "dev", min(4000, len(dev)), dict(per_src))