MS-MARCO-0G-120M

A long-context retrieval and question-answering benchmark with 1,143,371 documents and a repaired 192-question default evaluation (128 + 64), including 12 questions about 0G. The original 10,000-question construction pool (including 50 0G questions) is also retained. It extends the ms_100M bank from MSA-RAG-BENCHMARKS with 20 million additional text tokens and 655 new questions grounded in the added documents.

120M is the benchmark's nominal size. The actual corpus contains 125,708,601 raw tokens under the MSA-4B tokenizer: 105,708,601 original tokens plus exactly 20,000,000 added tokens. Counts exclude prompts, wrappers, and special tokens. This is a derived benchmark, not an official Microsoft MS MARCO release.

Repaired 192-question evaluation (v1.1.0)

The default questions configuration contains the frozen repaired 128 + 64 panel. It uses the complete 1,143,371-document 120M corpus. This release does not use the separate MS-MARCO-1.0 benchmark, its corpus, or its evaluation results. The document collection is unchanged.

Component	Questions	Role
Retained original regression questions	123	Original 128 minus five removed items
Replacement general questions	5	Complete the repaired 128
Previously selected development panel	64	52 general + 12 0G questions
Default evaluation	192	123 baseline + 57 general + 12 0G

All ordinals below are zero-based positions in the original 10,000-question pool. Removed regression ordinals: 1318, 3541, 3550, 6633, 8844. Replacement ordinals: 9434, 9489, 9703, 9824, 9840. Answers and/or canonical evidence were corrected at ordinals 3313, 4769, 6095, 7407; the exact old/new answers and evidence-change flags are recorded in evaluation/selection.json. The repaired evidence for question 3313 includes document ordinal 1034383, which belongs to the extended 120M bank.

Selection method: the 64 development questions and five replacements were selected using historical model scores, favoring the new_2500 checkpoint relative to flash_rag; they were not randomly sampled. This panel supports regression and development comparisons, not independent held-out generalization claims.

The original 10,000-question pool and its sidecars remain unchanged for provenance and compatibility; they do not incorporate the four corrected references. Use evaluation/ or the default questions configuration for the repaired benchmark.

Original construction pool

Component	Documents	Raw tokens	Questions
Original bank	961,686	105,708,601	9,345
Added MS MARCO passages, verbatim	172,465	19,015,686	—
Added synthetic general passages	9,088	970,222	605
Added synthetic 0G passages	132	14,092	50
Total	1,143,371	125,708,601	10,000

The original documents occupy positions 0:961686; the original questions occupy positions 0:9345. Their text, records, and ordering are preserved. The remaining 655 questions have cohort="general" (605) or cohort="0g" (50). The 0G questions cover Chain (14), Compute (13), Architecture (11), Storage (6), and DA (6).

The corpus split is the document collection.

Quick start

pip install datasets huggingface_hub

from datasets import load_dataset

repo = "0G-AI/MS-MARCO-0G-120M"
documents = load_dataset(repo, "documents", split="corpus")
questions = load_dataset(repo, "questions", split="test")

new_questions = questions.filter(lambda q: q["cohort"] != "baseline")
zero_g_questions = questions.filter(lambda q: q["cohort"] == "0g")

print(len(documents), len(questions), len(new_questions), len(zero_g_questions))
# 1143371 192 69 12
print(new_questions[0]["query"])

# Optional: unchanged construction pool, not the repaired evaluation.
question_pool = load_dataset(repo, "question_pool", split="pool")
assert len(question_pool) == 10000

Use streaming=True to iterate without materializing a dataset. For reproducible results, pass a specific repository commit as revision=.

Files

MS-MARCO-0G-120M/
├── README.md
├── TERMS.md
├── manifest.json
├── checksums.sha256
├── validation.json
├── documents.parquet
├── documents.jsonl
├── questions.parquet
├── questions.jsonl
├── evaluation/
│   ├── questions.parquet
│   ├── questions.jsonl
│   ├── qdata_msmarco_120M_192.pkl
│   ├── reference_positions.jsonl
│   └── selection.json
├── verify_evaluation.py
├── data/ms_120M/
│   ├── mdata_msmarco_120M.pkl
│   └── qdata_msmarco_120M.pkl
├── metadata/
│   ├── documents.jsonl
│   ├── questions.jsonl
│   └── tokenizer.json
├── sidecars/
│   ├── qrels.jsonl
│   ├── reference_positions.jsonl
│   └── new_evidence_sets.jsonl
└── verify.py

Within each question set, Parquet, JSONL, and pickle contain the same question content and ordering. The default repaired set is under evaluation/; root-level question files and data/ms_120M/qdata_msmarco_120M.pkl contain the unchanged 10,000-question pool.

Documents

documents.parquet and documents.jsonl have the same fields:

Field	Type	Meaning
id	string	Stable document identifier; treat it as an opaque key.
ordinal	integer	Zero-based position in the complete document bank.
text	string	Document text, without normalization or added wrappers.
text_sha256	string	SHA-256 of the UTF-8 text.
cohort	string	baseline, distractor, general, or 0g.

baseline identifies original documents, distractor identifies verbatim added MS MARCO passages, and general/0g identify synthetic added passages. "Distractor" describes the construction role; it does not guarantee irrelevance to every question.

Repaired evaluation questions

evaluation/questions.parquet and evaluation/questions.jsonl both use flat records: id, ordinal, evaluation_index, cohort, selection_group, query, answer, and reference_list. ordinal preserves the original pool position; evaluation_index is the contiguous 0–191 row index. selection_group is retained_regression, replacement, or original64. The frozen order is 123 retained regression questions, the original 64, then five replacements; filter on selection_group to recover the repaired 128.

evaluation/reference_positions.jsonl contains each question ID, original question ordinal, and every exact corpus position matching each canonical evidence text. evaluation/selection.json records composition, corrections, selection limitations, and source-artifact hashes. Evidence is for evaluation and must not be supplied as extra query input to retrieval models.

Original pool questions

Root-level questions.parquet exposes these columns directly:

Field	Type	Meaning
id	string	Stable question identifier.
ordinal	integer	Zero-based question position.
cohort	string	baseline, general, or 0g.
query	string	Question text.
answer	string	Reference answer.
reference_list	list of strings	Exact document texts used as canonical evidence.

In questions.jsonl, query, answer, and reference_list are nested under record; id, ordinal, and cohort remain top-level fields. This retains the original JSONL representation. Evidence is supplied for evaluation; it should not be given to a retrieval model as additional query input.

Original pool reference relationships

• sidecars/reference_positions.jsonl: one row per question, with question_ordinal and reference_positions. The latter is a list of lists: for each text in reference_list, it gives all matching document ordinals. Original documents can contain repeated text.
• sidecars/qrels.jsonl: question/document ID pairs. Original pairs have relation="inherited_reference" and sufficiency_revalidated=false. New pairs have relation="verified_sufficient_evidence_member", a canonical flag, and evidence-group entries in sets. These are structured JSON relationships, not a standard TREC TSV.
• sidecars/new_evidence_sets.jsonl: 655 rows with question_id, canonical_doc_ids, and alternative evidence_sets. Each set contains doc_ids, sufficient, and available verification information. A member of a sufficient multi-document set need not be sufficient on its own. flash and reader identify the construction judges, not human reviewers.

Source metadata

metadata/documents.jsonl contains one row for each of the 181,685 appended documents, keyed by id. It records the provenance class, cohort, token count, source coordinates, and available topic/generation model information. MS MARCO source coordinates include the cache revision, Arrow shard, row and passage indices, query ID, and source URL. 0G coordinates include the repository, commit, documentation path, and character offsets. Offsets use Unicode code points and refer to the source documentation, not the generated passage.

For verbatim MS MARCO passages, found_answering_a_frozen_question retains the construction screen's flag for a non-equivalent answer to an original question. false means not detected by that screen, not proven absent. Synthetic rows omit this field; their omission is not a negative label. Per-document source URLs for the original bank were not supplied by the inherited export.

metadata/questions.jsonl contains 655 rows with new-question topics, answer types, available answer aliases, and canonical evidence IDs. Aliases are auxiliary metadata; the original answer field remains the reference answer for comparable scoring.

metadata/tokenizer.json records the tokenizer file hashes, library versions used during construction, and counting rule. Token totals depend on that tokenizer.

Compatibility with MSA evaluation

from huggingface_hub import snapshot_download
from pathlib import Path
import pickle

root = Path(snapshot_download(
    "0G-AI/MS-MARCO-0G-120M",
    repo_type="dataset",
    allow_patterns=["data/ms_120M/mdata_msmarco_120M.pkl",
                    "evaluation/qdata_msmarco_120M_192.pkl"],
))
with (root / "data/ms_120M/mdata_msmarco_120M.pkl").open("rb") as f:
    memory = pickle.load(f)  # list[str], length 1,143,371
with (root / "evaluation/qdata_msmarco_120M_192.pkl").open("rb") as f:
    queries = pickle.load(f)  # list[dict], length 192

# Each query has query, answer, and reference_list.
assert len(queries) == 192

Only load pickle files from a trusted source. Parquet and JSONL are available for readers that do not use pickle.

Construction

The existing bank and question records were frozen. Verbatim MS MARCO v2.1 passages were selected after exact/normalized deduplication and MinHash screening. DeepSeek-V4-Flash-0731 generated additional source-grounded passages and questions, with Qwen3.8-27B used for independent reading and consistency checks. Retrieval-based conflict screening used Qwen3-Embedding-4B and Qwen3-Reranker-4B. Whole documents were selected to reach exactly 20,000,000 added raw tokens.

The added MS MARCO passages comprise 133,674 train, 14,831 validation, and 23,960 test passages from the upstream v2.1 cache. These source partitions are separate from this benchmark's evaluation split. The 0G material is grounded in 0gfoundation/0g-doc at commit b9e59f4a424990da35d86087db144c4d63e0b5b3; it describes that snapshot, not necessarily current 0G behavior.

Evaluation and limitations

Evaluate retrieval against the complete 1,143,371-document bank, including when scoring only the new or 0G questions. For the repaired panel, report baseline (123), general (57), and 0G (12) results separately, and disclose the selection method above. For optional full-pool evaluations, report original (9,345), new general (605), and 0G (50) results separately, and state your answer-normalization, retrieval, generation, and judge settings. Original pool questions and evidence are preserved; the repaired panel applies its four documented corrections. Adding documents can change retrieval difficulty and answer ambiguity; scores on the original bank and this expanded bank are not interchangeable.

Construction checks are model-based and mostly retrieval-limited, not exhaustive proof of answer uniqueness. The construction audit reported 23,860 added documents with non-equivalent answers to original questions. A 400-question sample contained 57 apparent new conflicts; checking those against the complete original bank reproduced 45 there, left 10 not reproduced, and 2 undecided. The 10/400 result suggests possible added ambiguity; it is not an enumerated correction set. The full original-question audit also left 103 questions undecided and 3 with incomplete candidate coverage.

Joint reading checked potentially complementary evidence for 134 new questions. Six used overlapping groups that covered every candidate pair, but not every larger combination. Two 0G questions retain sufficient evidence sets verified by a single reader after invalid references were removed. Some construction audits were performed against intermediate candidate corpora. This release verifies content and reference integrity; it does not claim that all semantic checks were rerun on the final corpus or that the bank is conflict-free.

The inherited bank is a benchmark-specific text collection, not an identified official MS MARCO corpus. Its per-document origins are incomplete. Appended verbatim passages and original chunks also differ in their text distribution. Synthetic questions and judge decisions can contain errors. The 50-question 0G subset is small. Because upstream test passages are included, this corpus should not be treated as held-out data for official MS MARCO leaderboard comparisons.

Integrity

The ordered corpus SHA-256 is:

69325b24d552440d9a7405eac09c45d9be33ed6e24b49cae762c9fba3375283c

manifest.json records counts, source identities, and data-file SHA-256 hashes. checksums.sha256 covers the release files. validation.json records the distribution checks, including the repaired evaluation. The original pool count remains question_count=10000 in the manifest; evaluation.question_count=192 identifies the default benchmark. To verify a complete download:

sha256sum -c checksums.sha256
pip install pyarrow
python verify.py .
# Or check only the repaired panel and its evidence against the memory bank:
python verify_evaluation.py .

verify.py verifies file hashes, document identities and ordering, JSONL/Parquet/pickle agreement, question references, evidence-set membership, and stored append-token totals. It also validates the repaired 192-question panel and its exact evidence positions. It does not run model-based semantic audits or retokenize the corpus.

Sources and terms

This benchmark incorporates MSA-RAG-BENCHMARKS, MS MARCO v2.1, and 0G documentation. It is intended for non-commercial research. Source-specific terms and third-party rights continue to apply; see TERMS.md. This release does not assign a blanket MIT or CC0 license to the source text.