The dataset viewer is not available for this dataset.
Error code: ConfigNamesError
Exception: AttributeError
Message: 'str' object has no attribute 'items'
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
config_names = get_dataset_config_names(
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
dataset_module = dataset_module_factory(
^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1217, in dataset_module_factory
raise e1 from None
File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1192, in dataset_module_factory
).get_module()
^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 700, in get_module
config_name: DatasetInfo.from_dict(dataset_info_dict)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/info.py", line 284, in from_dict
return cls(**{k: v for k, v in dataset_info_dict.items() if k in field_names})
^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'items'Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
MoC-RAG Benchmark: Typed Context Routing for Agentic Memory
A benchmark for evaluating whether routed, typed context experts (Mixture-of-Contexts RAG) improve retrieval and answer quality compared with flat RAG, under a fixed token budget.
Version 0.1.0. This dataset accompanies the paper
Matrix Context: Mixture-of-Contexts RAG for Robust and Inspectable Agent Memory
(10.5281/zenodo.20560139).
Why this benchmark
Flat RAG embeds everything into one index and retrieves nearest chunks for every query. MoC-RAG instead asks which typed context experts should be searched first, retrieves inside them, and assembles a token-budgeted, explainable pack. This dataset is built to test that claim where it should matter most: in the presence of hard negatives that flat RAG is tempted by but a router should avoid.
Contents
| file | rows | description |
|---|---|---|
contexts.jsonl |
1000 | typed context items (the memory store) |
queries.jsonl |
600 | queries / tasks |
gold.jsonl |
600 | gold, acceptable, and distractor labels |
splits/* |
— | train / validation / test query splits |
- 8 experts: user_memory, project_memory, document_rag, code_context, session_memory, decision_memory, policy_rules, tool_history
- 10 context types: fact, preference, goal, decision, rule, document, code, episode, summary, tool_result, profile
- 6 domains: project_architecture, user_persona, code_context, policy_rules, tool_session, document_rag
Hard negatives (5 kinds)
Each gold fact is surrounded by:
same_keyword_wrong_expert— same salient keyword, different typed expertsame_expert_wrong_scope— a true fact, but for a different project/personaoutdated_decision— a superseded decision (low confidence, old timestamp)contradictory_memory— a note asserting a conflicting valuestale_session_note— an old session mention with nothing decided
Query variants & robustness splits
Every gold topic is phrased five ways (the variant field on each query),
holding the gold label constant while varying lexical overlap and intent:
direct (keyword-aligned) · paraphrased · underspecified · cross_expert ·
adversarial (embeds a misleading term that lexically matches the contradictory
hard negative).
Beyond train / validation / test, three parallel test splits phrase the
same test topics three ways so robustness to lexical noise can be measured
directly: test_keyword, test_paraphrased, test_adversarial. A retriever
that scores well on test_keyword but degrades on test_adversarial is brittle
to paraphrase and misleading keywords.
Schema
contexts.jsonl
{"context_id": "ctx_000000", "expert": "decision_memory", "type": "decision",
"scope": "project:matrix-context", "content": "...", "tags": ["..."],
"importance": 0.95, "confidence": 0.92, "source": "synthetic_gold",
"created_at": "2026-06-04T10:00:00Z", "role": "gold"}
queries.jsonl
{"query_id": "q_000000", "query": "...", "task_type": "architecture_recall",
"expected_experts": ["decision_memory", "project_memory"],
"scope": "project:matrix-context", "difficulty": "easy", "domain": "..."}
gold.jsonl
{"query_id": "q_000000", "gold_context_ids": ["ctx_000000"],
"acceptable_context_ids": ["ctx_000001"],
"distractor_context_ids": ["ctx_000002", "..."],
"gold_answer": "SQLite", "gold_citations": ["ctx_000000"]}
Intended use
Evaluate retrieval + context efficiency + answer quality for agentic memory.
Report Recall@K, Precision@K, MRR, nDCG, distractor and token counts, useful
context ratio, expert routing accuracy, and (optionally) grounded answer
quality. Compare flat / BM25 / hybrid / metadata-filtered / reranked RAG against
MoC-RAG with top_experts in {1, 2, 3, all}.
Limitations and bias
The dataset is synthetic (template-generated, deterministic) and English only. It is designed to exercise the routing mechanism and tooling rigorously, not to make a strong general claim about real corpora; the roadmap is to grow it toward human-reviewed, real long-horizon memory. Synthetic facts about named projects are illustrative, not authoritative.
License
Apache-2.0.
Conformance
The reference implementation passes the MoC Contract v1 conformance suite and
is MoC API v1 Compatible / MoC Inspect v1 Compatible. See
moc_contract/ in the software repository and run
python -m moc_contract.conformance.
Citation
This benchmark accompanies the paper — please cite it:
Magaña Vsevolodovna, R. I. (2026). Matrix Context: Mixture-of-Contexts RAG for Robust and Inspectable Agent Memory. Zenodo. https://doi.org/10.5281/zenodo.20560139
@misc{magana2026matrixcontext,
title = {Matrix Context: Mixture-of-Contexts RAG for Robust and Inspectable Agent Memory},
author = {Magaña Vsevolodovna, Ruslan Idelfonso},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20560139},
url = {https://doi.org/10.5281/zenodo.20560139}
}
Software (Apache-2.0): github.com/agent-matrix/matrix-context ·
Leaderboard: huggingface.co/spaces/ruslanmv/moc-rag-leaderboard.
- Downloads last month
- 71