Kyle-han/infoseek-fact-store

# InfoSeek Fact Store

InfoSeek train/val questions에서 Wikidata claims를 이용해 복원한 FactHead 학습용 pseudo-supervision 데이터다.

이 데이터는 official InfoSeek evidence가 아니다. Gold QID, answer value/surface, Wikidata claims, property/unit/qualifier rules를 이용해 만든 positive fact와 hard negative fact다.

1. Layout

fact_store/
  data/facts/
    infoseek_numeric_facts_enriched.jsonl
    numeric_qid_to_fact_ids.json
    numeric_pid_to_fact_ids.json
    numeric_label_cache.json

    infoseek_string_facts.jsonl
    string_qid_to_fact_ids.json
    string_pid_to_fact_ids.json
    string_label_alias_cache.json

  data/maps/
    infoseek_train_numeric_question_fact_map_enriched_v2.jsonl
    infoseek_val_numeric_question_fact_map_enriched_v2.jsonl
    infoseek_train_string_question_fact_map_enriched_v2.jsonl
    infoseek_val_string_question_fact_map_enriched_v2.jsonl

    # raw string maps are kept for traceability
    infoseek_train_string_question_fact_map.jsonl
    infoseek_val_string_question_fact_map.jsonl

  data/docs/
    fact_coverage.md
    numeric_fact_enrichment.md
    match_vector_sampling_policy.md
    *_numeric_enrichment_summary_v2.json
    *_match_vector_eval.json
    infoseek_string_fact_summary.json
    infoseek_string_union_stats.json
    infoseek_string_enriched_v2_summary.json

  samples/
    sample_*_20.jsonl

2. Unified Training Interface

Both numeric/time and string/item enriched v2 maps expose the same training-facing fields:

selected_positive_fact_ids
positive_status_v2
hard_negative_fact_ids_by_type
hard_negative_meta_v2
fact_map_schema_version or numeric v2 fields
fact_map_kind when available

Use rows only if:

positive_status_v2 == "selected_positive"
selected_positive_fact_ids is not empty

Numeric/time and string/item still differ semantically:

numeric/time:
  hard_negative_meta_v2 has match_vector_key, match_vector_label, negative_tier
  match vectors encode entity/property/qualifier/value agreement

string/item:
  hard_negative_fact_ids_by_type contains negative ids
  hard_negative_meta_v2 is an empty dict in compact v2
  match vectors are not provided for string/item

3. Core Files

Numeric/time fact store:

data/facts/infoseek_numeric_facts_enriched.jsonl
data/maps/infoseek_train_numeric_question_fact_map_enriched_v2.jsonl
data/maps/infoseek_val_numeric_question_fact_map_enriched_v2.jsonl

String/item fact store:

data/facts/infoseek_string_facts.jsonl
data/maps/infoseek_train_string_question_fact_map_enriched_v2.jsonl
data/maps/infoseek_val_string_question_fact_map_enriched_v2.jsonl

Raw string maps are also kept:

data/maps/infoseek_train_string_question_fact_map.jsonl
data/maps/infoseek_val_string_question_fact_map.jsonl

Use train/val only for supervised fact recovery. Test/human are not included because the local files do not provide reliable gold QID mappings.

4. Negative Metadata

Numeric/time match vector examples:

11n0 = same entity, same property, qualifier not applicable, wrong value
10n1 = same entity, different property, qualifier not applicable, same value
01n0 = different entity, same property, qualifier not applicable, wrong value

Do not use these numeric/time match vectors as negatives:

1111
11n1

String/item negative groups:

same_entity_different_string_item_fact:
  meaning = same entity, different string/item fact

same_property_different_subject_value:
  meaning = same property, different subject/value

5. Fact Text

Both numeric/time and string/item fact files include text_for_encoder.

Use this as the simplest first input to a fact encoder:

z_fact = normalize(TextEncoder(fact["text_for_encoder"]))

Numeric example:

Subject: Rhine Falls. Property: height. Value: 23 metre.

String/item example:

Giant panda [P] endemic to [O] People's Republic of China

6. Minimal Unified Dataloader Sketch

import json
from pathlib import Path

root = Path("fact_store")

def load_jsonl(path):
    with open(path, encoding="utf-8") as f:
        for line in f:
            if line.strip():
                yield json.loads(line)

fact_by_kind = {
    "numeric_time": {
        row["fact_id"]: row
        for row in load_jsonl(root / "data/facts/infoseek_numeric_facts_enriched.jsonl")
    },
    "string_item": {
        row["fact_id"]: row
        for row in load_jsonl(root / "data/facts/infoseek_string_facts.jsonl")
    },
}

def iter_fact_rows(kind="numeric_time", split="train"):
    if kind == "numeric_time":
        path = root / "data/maps" / f"infoseek_{split}_numeric_question_fact_map_enriched_v2.jsonl"
    elif kind == "string_item":
        path = root / "data/maps" / f"infoseek_{split}_string_question_fact_map_enriched_v2.jsonl"
    else:
        raise ValueError(kind)

    fact_by_id = fact_by_kind[kind]
    for row in load_jsonl(path):
        pos = row.get("selected_positive_fact_ids", [])
        if row.get("positive_status_v2") != "selected_positive" or not pos:
            continue

        neg_groups = row.get("hard_negative_fact_ids_by_type", {})
        neg_ids = []
        for ids in neg_groups.values():
            neg_ids.extend(ids or [])

        neg_meta = row.get("hard_negative_meta_v2", {})
        if neg_meta:
            neg_ids = [
                fid for fid, meta in neg_meta.items()
                if meta.get("use_as_negative", True)
                and meta.get("match_vector_key") not in {"1111", "11n1"}
            ]

        yield {
            "kind": kind,
            "row": row,
            "positive_facts": [fact_by_id[fid] for fid in pos if fid in fact_by_id],
            "negative_facts": [fact_by_id[fid] for fid in neg_ids if fid in fact_by_id],
            "negative_meta": {fid: neg_meta[fid] for fid in neg_ids if fid in neg_meta},
        }

7. Suggested Training Setup

The intended module is a dual-encoder style FactHead.

query side:
  z_q_fact(e_i) = FactQueryHead(image, question, candidate_entity)

fact side:
  z_fact = FactItemHead(fact_text_or_fields)

training:
  pull z_q_fact close to positive facts
  push z_q_fact away from hard negative facts

For the first experiment, use oracle/gold entity from InfoSeek with-KB:

image + question + gold entity -> z_q_fact
selected_positive_fact_ids -> positives
hard_negative_fact_ids_by_type -> negatives

8. Current Summary

Numeric/time v2:

train rows = 190,983
train usable selected_positive rows = 140,772
val rows = 19,301
val usable selected_positive rows = 18,128

String/item enriched v2 (after full label-cache completion, 2026-07-04):

train rows = 743,065
train usable selected_positive rows = 579,619
train selected positive fact ids = 892,349
train hard negative fact ids = 22,903,608

val rows = 54,319
val usable selected_positive rows = 43,589
val selected positive fact ids = 62,849
val hard negative fact ids = 1,687,811

String-like positive coverage, local KG OR full Wikidata string/item:

train union positive = 586,193 / 743,065 = 78.89%
val union positive = 43,982 / 54,319 = 80.97%

Note: earlier versions reported 54.86% / 51.19%. The gap was mostly missing object/property labels in the label cache (21,835 ids unfetched due to API rate limits), not a structural ceiling. Completing the cache recovered +24-30pt. Remaining misses: no_match 126,906 (17.1%, partly recoverable via redirect-id mapping), no_string_answer_surface 35,729 (4.8%, not recoverable).

See:

data/docs/fact_coverage.md
data/docs/infoseek_string_fact_summary.json
data/docs/infoseek_string_union_stats.json
data/docs/infoseek_string_enriched_v2_summary.json
data/docs/*numeric_enrichment_summary_v2.json
data/docs/*match_vector_eval.json

9. Not Included

Image pixels are not included. Rows reference InfoSeek/OVEN image ids only. Training code must resolve image files separately.

Text evidence retrieval is not included here. This package only contains Wikidata fact supervision.