S1MB — System One Mosaic Benchmark
88 stored subsets · 13,443 cases · 27,303 decisions · test only
The default evaluation manifest contains 84 subsets · 11,843 cases · 24,103 decisions. 4 retrieval subsets are retained separately pending task/rubric compatibility review. This is an evaluation dataset, never a training pool.
Sources and traceability
All original dataset references, revisions, declared licenses and papers are indexed per subset. sources.json retains the full reference records. S1MB-original synthetic data records its author and content versions rather than claiming an external dataset source. Missing licenses or revisions remain explicitly unknown; inherited license metadata is not a new legal audit. Dataset-wrapper licenses may not cover the underlying corpora.
This evaluation dataset does not guarantee non-overlap with training data. Per-row provenance records original IDs, source splits, revisions and query/document references. manifest.json lists sampling membership including quarantined subsets. Use training-manifest.json for active evaluation; its train list is empty. Specify dataset version and membership when comparing results.
Task definitions and retrieval Score / Noul
Generic system prompts are empty. Task-specific instructions and criterion definitions follow the same task catalog as Bekko across splits. Instruction describes the task; state supplies the context (including a retrieval query and document); criteria define the typed answers. A document is not itself a Score criterion. Noul keeps explicit false/true definitions, Choice keeps distinct alternatives, and Score uses numeric rubric values.
Synthetic retrieval uses two independent decisions per case and a shared 20-pattern policy catalog: broad (7), standard (7), strict (6). Original teacher grades are recovered first. Broad/standard source positives receive max(original grade, 3); strict uses original grades. Negatives are not floored.
relevance_score: five ordered levels with values 0, 1, 2, 3, 4. The target is the policy-grade distribution. Expected Score is sum(p[i] * value[i]); fractional outputs are valid. Normalize by 4 only for normalized Score metrics.relevance_noul: retain versus do not retain under the policy. Target P(true) is the policy-defined expected utility u, and P(false)=1-u. The Noul instruction explicitly states the policy-specific numeric acceptance anchors. Anchors retain endpoints 0/1 and intermediate variation on a 0.05 grid, fixed per group/query. These are assigned synthetic probabilities, not measured or calibrated acceptance frequencies.
For example, a grade-3 target with a 0.85 utility anchor becomes Score level 3 and Noul P(true)=0.85. Do not impose P(true)=Score/4. Score criteria are stored in numeric order; Noul criteria use false/true. Other Score tasks retain their original level counts and values. Report the two tasks separately; each retrieval case contributes two decisions.
Labels are derived from source grades and policy anchors, not independently reannotated under every policy. Source-positive errors may survive broad/standard correction. Non-retrieval targets and authored generalization rubrics retain their source semantics. Do not reuse old predictions by case ID: compare input hashes and content versions. The actual effect on model generalization has not been measured.
S1MB generalization benchmarks
These six S1MB-original English synthetic evaluation subsets were created using GPT-6-Astra. The model authored the instructions, contexts, criteria, intended answers and rationales. Each subset contains 100 cases and 100 decisions, for 600 cases in total. They are evaluation data, not training data.
| Subset | Cases | Design |
|---|---|---|
s1mb-generalization-diverse-score |
100 | Diverse instructions, contexts and ordered scoring rubrics |
s1mb-generalization-diverse-noul |
100 | Individual free-form false/true definitions for each question |
s1mb-generalization-diverse-choice |
100 | Varied selection tasks with 2–8 alternatives |
s1mb-generalization-contextual-score |
100 | 20 fixed questions × 5 contexts; fixed rubric within each family |
s1mb-generalization-contextual-noul |
100 | 20 fixed questions × 5 cases; both context and false/true definitions vary |
s1mb-generalization-contextual-choice |
100 | 20 fixed questions × 5 contexts; fixed options and order within each family |
The purpose is to evaluate one aspect of generalization: adaptation to diverse instructions, contexts and decision criteria. The name does not guarantee non-overlap with a model's training data or establish a particular out-of-distribution evaluation protocol. Labels are author-intended and self-reviewed, not independently human-validated gold.
States can be plain text, JSON objects or arrays. Noul's false/true definitions specify the actual categories: true may denote a defect, denial or unresolved issue rather than approval. Both Noul subsets have 50 false and 50 true targets, with 200 distinct false descriptions and 200 distinct true descriptions across them. Contextual Noul tests context and criteria interpretation together, not an isolated effect of state alone.
S1MB generalization Score uses ordered numeric values in the 0–1 range. The diverse set uses 4, 5, 6 or 10 levels. Each contextual Score family uses 0, 0.25, 0.5, 0.75 and 1 once as its intended answer across five contexts. Choice uses 2–8 options with one intended correct answer. Each contextual Choice family uses as many distinct intended answers as its option count and five cases allow.
Political and religious cases use explicitly stated fictional commitments and circumstances; identities alone do not determine answers. Cases also cover consent, counterfactuals, negation, quantifiers, sarcasm, boundaries, accessibility and operational decisions.
Each contextual group_id identifies a five-case question family. Keep its cases together in any derived split. Report task-level metrics and optionally the proportion of families with all five top-criterion answers correct. Five related cases are not five independent scenarios. For diverse subsets, group IDs retain related shared states.
Full questions, criteria, targets and rationales are available in generalization/<subset>/REVIEW.md. Those review documents contain supervision and must not be sent to a model. Subset metadata records authorship and content versions; match predictions to input hashes.
Format and loading
The Arrow schema is system_one.v1:
input.state_jsoncontains the shared context as encoded JSON.input.decisionscontains typed questions, instructions and criteria, or explicit document-ranking problems.targetscontains labels and probability distributions.- Case/group IDs and provenance support tracking and evaluation, not model input.
Instructions and descriptions preserve their structured JSON values at the logical interface. Use the schema adapter to obtain model inputs:
from datasets import load_from_disk
from bekko_system_one.dataset_schema import inference_input, to_jev_request
rows = load_from_disk("<release>/s1mb-generalization-contextual-noul")["test"]
model_input = inference_input(rows[0])
jev_request = to_jev_request(rows[0])
targets = rows[0]["targets"]
to_jev_request passes Noul's individual false/true definitions to the API. Do not substitute generic yes/no descriptions. Labels, rationales and provenance must remain outside inference inputs.
Evaluation semantics
- Noul probabilities are over false/true categories; use binary Brier score and classification metrics.
- Choice probabilities are over competing options; keep IDs aligned with descriptions and targets.
- Score probabilities are over ordered rubric levels. Compute expected scores using criterion values and normalize errors by the scale range. Do not equate option indices with numeric values.
- Pointwise relevance is an independent Score per document, not a probability distribution across documents.
- Ranking distributions and pairwise preferences retain their own semantics. Unjudged or mined-negative documents are not automatically absolute false labels.
- Preserve soft targets, nested inputs and linked groups. For model comparisons, specify rendering, token budgets, candidate order, tie handling and aggregation weights.
Release files and verification
training-manifest.json: active evaluation subsets; no train entries.all-data-manifest.json: all stored subsets;quarantine-manifest.json: excluded subsets and reasons.manifest.json/sampling-exceptions.json: sampling membership and exceptions.dataset_metadata.json,sources.json,SOURCES.md: descriptions, counts, original references and declared licenses.task-catalog.json/score-policy-catalog.json: exact task rules, policy patterns, numeric anchors and seed.schema.json: system_one.v1;verification.json: structural checks and transformation counts, not model scores._renewal/receipts.json: source/output shard hashes and persisted logical hashes;_renewal/reproduction/: builder snapshots.checksums.json: hashes of all packaged files except itself. Generalization review files contain answers and are not inference input.
Runtime loading uses Hugging Face datasets and the Bekko System One schema/renderer. Source acquisition paths are provenance, not runtime dependencies. Validate changes in a separate output directory.
Stored subsets and active membership
| Subset | Cases | Decisions | Status | Original references |
|---|---|---|---|---|
| aegis2 | 100 | 148 | include_with_caveat | sources |
| ag_news | 100 | 100 | include | sources |
| aqua_rat | 100 | 100 | include | sources |
| arc | 100 | 100 | include_with_caveat | sources |
| arct | 100 | 100 | include | sources |
| argument_quality | 100 | 200 | include_with_caveat | sources |
| babi_nli | 100 | 100 | include | sources |
| banking77 | 100 | 100 | include_with_caveat | sources |
| bbq | 100 | 100 | include_with_caveat | sources |
| boardgameqa | 100 | 100 | include | sources |
| canttalk | 100 | 100 | include | sources |
| civil_comments | 100 | 700 | include_with_caveat | sources |
| cladder | 100 | 100 | include_with_caveat | sources |
| clinc | 151 | 151 | include_with_caveat | sources |
| contract_nli | 100 | 1700 | include | sources |
| corr2cause | 100 | 100 | include | sources |
| creak | 100 | 100 | include | sources |
| crows_pairs | 100 | 100 | include | sources |
| dbpedia | 100 | 100 | include_with_caveat | sources |
| defeasible_nli | 100 | 100 | include_with_caveat | sources |
| enron_spam | 100 | 100 | include | sources |
| esci | 100 | 100 | include | sources |
| ethics | 100 | 100 | include | sources |
| ethos | 100 | 100 | include_with_caveat | sources |
| few_nerd | 100 | 358 | include_with_caveat | sources |
| fol_nli | 100 | 100 | include | sources |
| followir_core17 | 100 | 200 | include_with_caveat | sources |
| followir_news21 | 100 | 200 | include_with_caveat | sources |
| followir_robust04 | 100 | 200 | include_with_caveat | sources |
| go_emotions | 100 | 2800 | include_with_caveat | sources |
| gretel_pii | 100 | 353 | include | sources |
| gsm8k | 100 | 100 | include | sources |
| hans | 100 | 100 | include | sources |
| hatecheck | 100 | 100 | include_with_caveat | sources |
| hh_rlhf | 100 | 100 | include_with_caveat | sources |
| hwu64 | 100 | 100 | include_with_caveat | sources |
| impli | 100 | 100 | include | sources |
| lexcomp | 100 | 100 | include | sources |
| logical_entailment | 100 | 100 | include | sources |
| lonli | 100 | 100 | include_with_caveat | sources |
| massive | 100 | 200 | include_with_caveat | sources |
| miqa | 100 | 100 | include_with_caveat | sources |
| mtop | 100 | 200 | include_with_caveat | sources |
| nlsat | 100 | 100 | include | sources |
| openbookqa | 100 | 100 | include | sources |
| patent_similarity | 100 | 100 | include | sources |
| paws | 100 | 100 | include_with_caveat | sources |
| plane | 100 | 100 | include | sources |
| poem_sentiment | 100 | 100 | include | sources |
| qasper | 100 | 100 | include | sources |
| quartz | 100 | 100 | include_with_caveat | sources |
| robust_lr | 100 | 100 | include_with_caveat | sources |
| ruletaker | 100 | 100 | include | sources |
| scicite | 100 | 100 | include | sources |
| scitail | 100 | 100 | include | sources |
| scone | 100 | 100 | include | sources |
| sdoh_nli | 100 | 100 | include_with_caveat | sources |
| sgd | 100 | 100 | include | sources |
| snli | 100 | 100 | include_with_caveat | sources |
| spartqa | 100 | 100 | include | sources |
| stepgame | 100 | 100 | include | sources |
| synthetic_relevance__nanobeir__NanoArguAna | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoClimateFEVER | 400 | 800 | quarantine | sources |
| synthetic_relevance__nanobeir__NanoDBPedia | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoFEVER | 400 | 800 | quarantine | sources |
| synthetic_relevance__nanobeir__NanoFiQA2018 | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoHotpotQA | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoMSMARCO | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoNFCorpus | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoNQ | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoQuoraRetrieval | 400 | 800 | quarantine | sources |
| synthetic_relevance__nanobeir__NanoSCIDOCS | 400 | 800 | include_with_caveat | sources |
| synthetic_relevance__nanobeir__NanoSciFact | 400 | 800 | quarantine | sources |
| synthetic_relevance__nanobeir__NanoTouche2020 | 392 | 784 | include_with_caveat | sources |
| synthetic_relevance__nanocoir | 800 | 1600 | include_with_caveat | sources |
| temporal_nli | 100 | 100 | include | sources |
| tracie | 100 | 100 | include | sources |
| typed_decisions | 100 | 500 | include | sources |
| ud_ewt | 100 | 1509 | include_with_caveat | sources |
| winobias | 100 | 100 | include | sources |
| winowhy | 100 | 100 | include | sources |
| wiqa | 100 | 100 | include | sources |
| s1mb-generalization-diverse-score | 100 | 100 | include_with_caveat | sources |
| s1mb-generalization-diverse-noul | 100 | 100 | include_with_caveat | sources |
| s1mb-generalization-diverse-choice | 100 | 100 | include_with_caveat | sources |
| s1mb-generalization-contextual-score | 100 | 100 | include_with_caveat | sources |
| s1mb-generalization-contextual-noul | 100 | 100 | include_with_caveat | sources |
| s1mb-generalization-contextual-choice | 100 | 100 | include_with_caveat | sources |
Quarantined retrieval subsets
synthetic_relevance__nanobeir__NanoClimateFEVER: Quarantine from the answer-evidence evaluation pool because an inspected source-positive sample directly conflicts with the rubric: The query claims Arctic sea-ice loss is matched by Antarctic sea-ice gain; the source-positive document is an encyclopedia overview of Antarctica that does not state this comparison. The teacher assigned 0, then the qrel-positive floor changed it to 3. Keep the saved raw teacher score and source label separately; do not replace the teacher instruction with source-task semantics while retaining the floored target. Evidence case: synthetic_relevance__nanobeir__NanoClimateFEVER:hakari-bench/NanoBEIR-en/NanoClimateFEVER:1575:doc-1.synthetic_relevance__nanobeir__NanoFEVER: Pending task/rubric compatibility review; this does not establish that the original paraphrase label or fact-verification qrel is wrong. Quarantine from the answer-evidence evaluation pool because an inspected source-positive sample raises an unresolved interpretation conflict with the answer-evidence rubric: The claim says Wolfgang Amadeus Mozart was a wrestler; the source-positive document is a Mozart biography with no wrestling evidence. The teacher assigned 0, then the qrel-positive floor changed it to 3. Keep the saved raw teacher score and source label separately; do not replace the teacher instruction with source-task semantics while retaining the floored target. Evidence case: synthetic_relevance__nanobeir__NanoFEVER:hakari-bench/NanoBEIR-en/NanoFEVER:34583:doc-8.synthetic_relevance__nanobeir__NanoQuoraRetrieval: Quarantine from the answer-evidence evaluation pool because an inspected source-positive sample directly conflicts with the rubric: The query asks what “being civilized” means; the source-positive duplicate question asks the same thing but contains no definition or answer. The teacher assigned 0, then the qrel-positive floor changed it to 3. Keep the saved raw teacher score and source label separately; do not replace the teacher instruction with source-task semantics while retaining the floored target. Evidence case: synthetic_relevance__nanobeir__NanoQuoraRetrieval:hakari-bench/NanoBEIR-en/NanoQuoraRetrieval:56198:doc-1.synthetic_relevance__nanobeir__NanoSciFact: Quarantine from the answer-evidence evaluation pool because an inspected source-positive sample directly conflicts with the rubric: The query claims insomnia can be treated with cognitive behavioral therapy; the source-positive document describes circadian-rhythm sleep-disorder symptoms, not CBT treatment efficacy. The teacher assigned 0, then the qrel-positive floor changed it to 3. Keep the saved raw teacher score and source label separately; do not replace the teacher instruction with source-task semantics while retaining the floored target. Evidence case: synthetic_relevance__nanobeir__NanoSciFact:hakari-bench/NanoBEIR-en/NanoSciFact:641:doc-1.