anonyMicky/anonymized-dataset

Anonymized Dataset

Review artifact prepared for anonymous submission. This release contains four question sets projected onto a fixed retrieval corpus, with final phase-2 document relevance judgments and inclusion/exclusion annotations.

Dataset	Questions	Qrels	Inclusion qrels	Exclusion qrels
BCP	80	10,968	9,717	1,251
TRQA	68	7,512	4,891	2,621
HotpotQA	204	10,033	9,210	823
DSQA	71	8,698	5,987	2,711

Layout

Each dataset has a separate directory and is a split of the configurations above.

• queries.jsonl: id, question, answer, canary.
• qrels.jsonl: query_id, iteration, doc_id, relevance, canary.
• qrels_inclusion.jsonl and qrels_exclusion.jsonl: same qrel schema, partitioned by role. Their disjoint union equals qrels.jsonl.
• qrel_roles.jsonl: the qrel schema plus role (inclusion or exclusion).
• DSQA additionally provides answer_sets.jsonl: id, encoded answer_set, canary, preserving complete members and accepted aliases for set evaluation.

HotpotQA's source stores each gold answer in a one-element list. The sole member is preserved as the scalar answer string to keep the original queries schema.

Inclusion evidence supports why an accepted candidate satisfies the question. Exclusion evidence rules out a plausible candidate or alternative interpretation. Evidence serving both functions is assigned to inclusion in the document-level partition. Roles do not change document relevance.

The qrels retain distinct document IDs for accepted duplicates. Each unique question–document pair occurs once in the combined qrels.

Encoding

Question IDs, questions, answers, and qrel document IDs are obfuscated with base64 over UTF-8 bytes XORed with a repeated SHA-256 digest of the row's canary. The DSQA answer_set string uses the same scheme and decodes to JSON. Encoded query IDs join directly to encoded qrel query IDs.

This is reversible contamination-prevention obfuscation, not encryption or access control. Repository privacy controls access.

import base64
import hashlib
import json

def decode_field(value, canary):
    key = hashlib.sha256(canary.encode("utf-8")).digest()
    raw = base64.b64decode(value)
    return bytes(c ^ key[i % len(key)] for i, c in enumerate(raw)).decode("utf-8")

# Decode a query row using its canary:
# question = decode_field(row["question"], row["canary"])
# members = json.loads(decode_field(set_row["answer_set"], set_row["canary"]))

deobfuscate.py decodes any configuration from a local JSONL file:

python3 deobfuscate.py bcp/queries.jsonl > queries_decoded.jsonl
python3 deobfuscate.py bcp/qrels.jsonl > qrels_decoded.jsonl

The corpus itself, construction logs, evaluation outputs, and private provenance are not redistributed in this artifact. Document IDs refer to the fixed corpus ordering used for the final relevance judgments.