Upload scripts/build_v15_mix.py with huggingface_hub

scripts/build_v15_mix.py

@@ -0,0 +1,87 @@
+import hashlib
+import json
+import random
+
+random.seed(15)
+
+
+def fingerprint(row):
+    state = row["state"] if isinstance(row["state"], str) else json.dumps(
+        row["state"], sort_keys=True)
+    questions = json.dumps(row["questions"], sort_keys=True)
+    return hashlib.md5((state + questions).encode()).hexdigest()
+
+
+def load(path):
+    try:
+        with open(path, encoding="utf-8") as f:
+            return [json.loads(line) for line in f]
+    except FileNotFoundError:
+        print("missing", path)
+        return []
+
+
+spec = [("data/adjud_train.jsonl", 16000, 1),
+        ("data/jb7_train.jsonl", 6000, 1),
+        ("data/jb6_train.jsonl", 5000, 1),
+        ("data/sft_v6.jsonl", 3000, 1),
+        ("data/sft_v8b.jsonl", 2000, 1),
+        ("data/hf4.jsonl", 4000, 1),
+        ("data/hf3b.jsonl", 3000, 1),
+        ("data/hf2b.jsonl", 3000, 1),
+        ("data/hf5_bb.jsonl", 1500, 1),
+        ("data/rl_sft_v9.jsonl", 1500, 1),
+        ("data/sft_v9.jsonl", 1500, 1)]
+
+seen = set()
+train = []
+for path, cap, copies in spec:
+    rows = load(path)
+    random.shuffle(rows)
+    kept = 0
+    for row in rows:
+        h = fingerprint(row)
+        if h in seen:
+            continue
+        seen.add(h)
+        row["id"] = f"v15:{path.split('/')[-1]}:{kept}:{h[:8]}"
+        train.append(row)
+        kept += 1
+        if kept >= cap:
+            break
+    print(path, kept)
+
+rendered = load("data/adjud_rendered.jsonl")
+unique_rendered = {}
+for row in rendered:
+    unique_rendered.setdefault(fingerprint(row), row)
+rendered = list(unique_rendered.values())
+random.shuffle(rendered)
+repeat = 8
+for rep in range(repeat):
+    for i, row in enumerate(rendered):
+        clone = dict(row)
+        clone["id"] = f"v15:render:{rep}:{i}:{fingerprint(row)[:8]}"
+        if clone.get("group_id"):
+            clone["group_id"] = f"{clone['group_id']}-v15r{rep}"
+        train.append(clone)
+print("rendered unique", len(rendered), "repeated", len(rendered) * repeat)
+random.shuffle(train)
+with open("data/v15_mix.jsonl", "w", encoding="utf-8") as f:
+    for row in train:
+        f.write(json.dumps(row, ensure_ascii=False) + "\n")
+
+train_fps = {fingerprint(row) for row in train}
+dev, dev_seen = [], set()
+for path in ("data/adjud_dev.jsonl", "data/jb6_dev.jsonl", "data/jb7_dev.jsonl"):
+    for row in load(path):
+        h = fingerprint(row)
+        if h in train_fps or h in dev_seen:
+            continue
+        dev_seen.add(h)
+        row["id"] = f"v15dev:{path.split('/')[-1]}:{len(dev)}:{h[:8]}"
+        dev.append(row)
+with open("data/v15_dev.jsonl", "w", encoding="utf-8") as f:
+    for row in dev:
+        f.write(json.dumps(row, ensure_ascii=False) + "\n")
+print("train", len(train), "dev", len(dev), "train/dev overlap", len(train_fps & dev_seen))