File size: 2,662 Bytes
dc6ba49 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | import hashlib, json
from collections import Counter
from alea.genbase import emit, q_choice, q_noul, q_score, validate_row, write_jsonl
def sig(r):
return hashlib.sha256((json.dumps(r["state"],sort_keys=True,ensure_ascii=False)
+ json.dumps(r["questions"],sort_keys=True,ensure_ascii=False)).encode()).hexdigest()
def convert(ex, idx):
kind = ex["kind"]
keys = [o["key"] for o in ex["options"]]
texts = [f'{o["key"]}: {o["description"]}' for o in ex["options"]]
gold = ex["answer_key"]
if gold not in keys:
return None
gi = keys.index(gold)
if kind == "choice":
q = q_choice("answer", ex["question"], texts,
[1.0 if i == gi else 0.0 for i in range(len(texts))])
elif kind == "score":
q = q_score("score", ex["question"], texts,
[1.0 if i == gi else 0.0 for i in range(len(texts))])
elif kind == "noul":
q = q_noul("answer", ex["question"], 1.0 if gold == "yes" else 0.0)
else:
return None
row = emit("tev1_" + ex["source"], idx, ex["state"], [q])
row["id"] = "tev1-" + ex["id"]
row["provenance"] = {"dataset": "togethercomputer/tev1", "source": ex["source"],
"group_id": ex["group_id"], "split": ex["split"],
"upstream": ex.get("provenance")}
validate_row(row)
return row
def load(path):
rows, seen = [], set()
stats = Counter()
for i, line in enumerate(open(path)):
ex = json.loads(line)
r = convert(ex, i)
if r is None:
stats["unsupported"] += 1
continue
h = sig(r)
if h in seen:
stats["dup"] += 1
continue
seen.add(h)
rows.append(r)
stats["src:" + ex["source"]] += 1
return rows, seen, stats
train, tr_sigs, tr_stats = load("/root/tev1/data/new-v1/records/train.jsonl")
dev, dv_sigs, dv_stats = load("/root/tev1/data/new-v1/records/dev.jsonl")
tr_groups = {r["provenance"]["group_id"] for r in train}
overlap = sum(1 for r in dev if r["provenance"]["group_id"] in tr_groups)
dev = [r for r in dev if sig(r) not in tr_sigs]
transfer, tf_sigs, tf_stats = load("/root/tev1/data/v2.1/records/transfer.jsonl")
transfer = [r for r in transfer if sig(r) not in tr_sigs | dv_sigs]
write_jsonl(train, "/root/alea/data/tev1_train.jsonl")
write_jsonl(dev, "/root/alea/data/tev1_dev.jsonl")
write_jsonl(transfer, "/root/alea/data/tev1_transfer.jsonl")
print("train", len(train), dict(tr_stats))
print("dev", len(dev), "group_overlap_rows_removed_only_if_dup:", overlap, dict(dv_stats))
print("transfer", len(transfer), dict(tf_stats))
|