SoroushVahidi/nl-optimization-instantiation-metrics

Natural-Language Optimization Instantiation Metrics (v1.2)

This dataset is a text-free metrics release for a retrieval-assisted natural-language optimization instantiation pipeline. It contains per-example and aggregate evaluation outcomes for a frozen pipeline evaluated on the NLP4LP benchmark, the OptMath external validation domain (added in v1.1), and, as of v1.2, a 13-method schema-retrieval/grounding diagnostic suite (per-query breakdowns, bottleneck taxonomy, threshold sensitivity, oracle-gap decomposition, baseline-vs-selective significance testing, and historical-compatibility bookkeeping) covering all 16 configs and 5,954 rows described below.

Soroush Vahidi created and releases the evaluation metrics and derived outcomes in this repository. NLP4LP is a third-party benchmark/resource maintained by the Udell Lab. Soroush Vahidi does not own NLP4LP or OptiMUS.

Raw NLP4LP problem text, gold formulations, gold structures, source IDs, prompts, completions, external example text, provider raw outputs, and local/private metadata are not included. Public rows use project-local opaque identifiers. Users who need the source benchmark must obtain NLP4LP directly from its canonical source under its own access and licensing terms: https://huggingface.co/datasets/udell-lab/NLP4LP

What Is Being Evaluated?

Many optimization problems are first written in natural language and later translated into mathematical models. This release evaluates a narrower intermediate task: natural-language optimization instantiation. Given a natural-language problem description, the evaluated system retrieves a compatible optimization schema from a fixed catalog and then fills scalar parameter slots from numeric evidence.

Schema retrieval means selecting the top-ranked optimization schema for a problem description. Instantiation readiness measures whether the retrieved and filled schema has enough correctly typed scalar information to be useful for downstream modeling.

The frozen method represented here is TF-IDF schema retrieval plus typed greedy scalar grounding with deterministic multiplicative ratio-word extraction. The patch adds deterministic extraction for expressions such as multiplicative ratio words, while leaving schema retrieval, slot choice, and evaluation thresholds unchanged.

Configs

per_query_metrics

One row represents one benchmark query evaluated by the frozen method. It does not include the native benchmark ID or text.

Fields:

• example_uid: project-local opaque identifier derived from the native source ID using a private keyed HMAC; the key and mapping are not released.
• variant: benchmark/query variant label from the source evaluation output. In this v1 release it is orig.
• method: method identifier. This is the source baseline column renamed for clarity; in this v1 release it is tfidf.
• schema_hit: whether the retrieved top-1 schema matched the gold schema.
• ordinary_ready: param_coverage >= 0.8 AND type_match >= 0.8.
• strict_ready: schema_hit AND param_coverage >= 0.8 AND type_match >= 0.8.
• n_expected_scalar: structural count of expected scalar slots used for evaluation. It is a small integer count and does not include slot names, variable names, formulation text, or problem text.
• n_filled: number of scalar slots filled by the evaluated method.
• param_coverage: scalar-slot coverage metric.
• type_match: scalar-slot type compatibility metric.
• exact5: exact-value metric at 5% tolerance where comparable; null means no comparable exact-value denominator for that row.
• exact20: exact-value metric at 20% tolerance where comparable; null means no comparable exact-value denominator for that row.
• key_overlap: schema/key overlap metric.

method_summary

Aggregate metrics for the pre-patch and final patched method variants. Counts are integer counts; rates and means are floating-point metrics.

strict_readiness_summary

Small aggregate table for strict-readiness counts and rate changes.

patch_transitions

Transition counts comparing pre-patch and patched outcomes. ID lists from the source transition file are deliberately excluded.

patch_case_metrics

Redacted per-case patch metrics for the changed-query audit. This config excludes source sentences, native query IDs, extracted mention strings, affected slot names, and multiplier expression text. It keeps only opaque IDs, change-status categories, readiness booleans, and numerical before/after metrics.

optmath_external_validation

An external validation configuration introduced in v1.1 containing 1,000 detailed query-level mathematical parameter extraction metrics based on the OptMath dataset (shushulei/OptMATH-Train). This acts as an out-of-domain generalization and robustness test beyond the core NLP4LP benchmark. To guarantee zero reconstruction risk and strict adherence to text-free metrics, it contains absolutely no plaintext, retrieved passages, model responses, exact parameter arrays, or raw problem descriptions.

Fields:

• example_uid: Stable project-local opaque identifier derived via HMAC from the native source ID; raw mappings are not published.
• evaluation_domain: Constant string (OptMATH).
• raw_ast_recall: Recall score computed directly over total abstract syntax tree constants.
• text_supported_recall: Recall score evaluated specifically over text-supported numeric slots.
• derived_recall: Recall score evaluated over slots requiring mathematical derivation.
• is_complete: Boolean indicator representing whether the instance was completely and perfectly extracted.
• num_gold_ast: Total count of abstract syntax tree numeric constants defined by the gold benchmark.
• num_matched_ast: Total count of correctly matched AST constants by the model.
• num_gold_text_supported: Count of expected text-supported constants.
• num_matched_text_supported: Count of matched text-supported constants.
• num_gold_derived: Count of expected derived constants.
• num_matched_derived: Count of matched derived constants.
• num_extracted: Total number of numeric values extracted by the model.
• num_unsupported: Count of extracted numbers unsupported by the source text annotations.

per_query_grounding_diagnostics (new in v1.2)

Per-query outcomes for all 13 evaluated schema-retrieval/grounding method variants (4,303 rows = 331 queries × 13 methods), extending per_query_metrics (which covers the single frozen tfidf method) to the full method comparison used by grounding_method_summary. Fields mirror per_query_metrics (schema_correct, ordinary_ready, strict_ready, param_coverage, type_match, exact5, exact20, key_overlap) plus predicted_example_uid/gold_example_uid (opaque, HMAC-derived — equal to example_uid on schema hits, different on schema misses) and false_ready, a boolean flagging cases that pass the readiness thresholds while schema_correct is false. No source text or native IDs are included, consistent with the rest of the release.

grounding_method_summary (new in v1.2)

One row per evaluated method (13 rows): aggregate schema_R1, ready counts/rates, Wilson 95% confidence intervals on strict-readiness (strict_wilson95_low/strict_wilson95_high), false_ready_count, and coverage/type-match/exact-match means. This is the method-comparison counterpart to method_summary (which only tracks the two pre-patch/patched tfidf states).

false_ready_details (new in v1.2)

The 124 rows across all 13 methods where false_ready is true in per_query_grounding_diagnostics (readiness thresholds passed despite an incorrect schema), with an added gold_scalar_count and a taxonomy label (fewer_or_easier_scalar_slots or structurally_similar_schema) categorizing why the false positive occurred. No mention strings, slot names, or source text are included — only opaque IDs, counts, and category labels.

bottleneck_taxonomy (new in v1.2)

77 rows (13 methods × up to 6 buckets) counting failure-mode categories per method: schema_wrong, schema_correct_coverage_failure, schema_correct_typematch_failure, schema_correct_both_coverage_and_typematch_failure, schema_correct_ready_but_exact20_not_full, schema_correct_ready_and_exact20_full. Complements grounding_method_summary by breaking each method's failure count down by cause.

threshold_sensitivity (new in v1.2)

27 rows sweeping coverage_threshold/type_match_threshold over {0.8, 0.9, 1.0} (a 3×3 grid) for the top 3 methods, reporting ordinary_ready_count/strict_ready_count/false_ready_count at each threshold pair. Shows how sensitive the readiness definition is to the two threshold choices used elsewhere in this release (0.8/0.8).

oracle_decomposition (new in v1.2)

A single-row decomposition of the gap between a retrieval oracle ceiling (retrieval_ceiling_schema_R1=1.0, i.e. perfect schema retrieval) and the achieved strict-ready counts for the baseline (tfidf_typed_greedy) and selective (tfidf_selective_grounding_rerank) methods, isolating how much of the remaining gap is attributable to retrieval error versus downstream grounding.

baseline_vs_selective_transition_details (new in v1.2)

Per-example transition records (12 rows, opaque example_uid/gold_example_uid only) for the queries where the baseline (tfidf_typed_greedy) and selective-rerank (tfidf_selective_grounding_rerank) methods disagree on strict_ready, ordinary_ready, or schema_correct — the discordant pairs underlying the McNemar tests in baseline_vs_selective_transitions.

baseline_vs_selective_transitions (new in v1.2)

Paired McNemar significance tests (3 rows, one per metric: strict_ready, ordinary_ready, schema_correct) comparing tfidf_typed_greedy against tfidf_selective_grounding_rerank, with discordant-pair counts, exact two-sided binomial p-values, and opaque discordant-example UID lists. mcnemar_note flags metrics with fewer than 10 discordant pairs as small-sample indications only.

historical_compatibility (new in v1.2)

Four rows recording pre-v1-pipeline historical readiness numbers for four methods, retained for internal compatibility bookkeeping only. Every row's note field explicitly warns these are a distinct, older measurement era and must not be mixed or compared with the fresh, canonical 257/331 and 331-query metrics reported elsewhere in this dataset. source intentionally does not name any internal file path or project directory — it identifies only that the numbers come from a historical, pre-release internal archive.

exact_metric_relationship (new in v1.2)

13 rows (one per method) clarifying how the exact5/exact20 exact-value metrics relate to readiness counts — in particular, that exact5/exact20 are only populated when schema_hit is true (schema-gated by construction), so they cannot be directly compared to the unconditional ordinary_ready_count/strict_ready_count columns without accounting for that gating.

Current Canonical Results

The final patched method has:

• schema R@1: 0.909366
• ordinary instantiation-ready: 265 / 331 = 0.800604
• strict instantiation-ready: 255 / 331 = 0.770393
• strict-ready patch gain: +8
• strict-ready losses: 0

These are point-in-time results for the frozen method state represented by the source repository at commit 6bb75a4c4bed02c458ac30b4af206a2802fce095.

Limitations

This is a benchmark-specific metrics release over NLP4LP orig evaluation queries. It does not contain the source benchmark examples, so users cannot inspect or resolve individual examples from this dataset alone. Users needing problem text or gold structures must obtain NLP4LP from the canonical upstream source.

The dataset includes remaining non-ready cases, schema retrieval failures, and type/coverage failures. Null exact5 and exact20 values indicate rows without a comparable exact-value denominator, not successful or failed exact matching.

The metrics are derived from one frozen pipeline/version. They should not be treated as claims about all natural-language optimization systems or all optimization-modeling benchmarks.

Version History

• v1.2 (current): additive-only release. Adds 10 new configs (per_query_grounding_diagnostics, grounding_method_summary, false_ready_details, bottleneck_taxonomy, threshold_sensitivity, oracle_decomposition, baseline_vs_selective_transition_details, baseline_vs_selective_transitions, historical_compatibility, exact_metric_relationship) covering a 13-method schema-retrieval/grounding comparison, bringing the release from 6 configs / 1,377 rows to 16 configs / 5,954 rows. No existing v1.1 config, column, or row was changed, renamed, or removed — code reading any v1.1 config (per_query_metrics, method_summary, strict_readiness_summary, patch_transitions, patch_case_metrics, optmath_external_validation) continues to work unchanged. size_categories corrected from n<1K to the factually accurate 1K<n<10K (true even at the prior v1.1 row count of 1,377).
• v1.1: added the optmath_external_validation config (1,000 rows, OptMATH out-of-domain validation).
• v1.0: initial release — per_query_metrics, method_summary, strict_readiness_summary, patch_transitions, patch_case_metrics (1,377 rows across 5 configs).

How This Dataset Differs from Existing Resources

NLP4LP provides benchmark problem resources. OptiMUS provides optimization-modeling systems and papers around NLP4LP. NL4Opt, ORLM, and PaMOP are related natural-language optimization or optimization-modeling resources. This release instead provides a text-free, downloadable metrics artifact: project-generated per-example evaluation outcomes, readiness indicators, and aggregate method metrics for a specific retrieval-assisted instantiation pipeline.

In the resources examined for this release, we did not identify an equivalent downloadable metrics-only artifact with this exact evaluation representation.

Relationship to Existing SoroushVahidi Datasets

This dataset is distinct from the existing original research datasets under SoroushVahidi:

• lafc-evict: cache-eviction/counterfactual supervision data.
• module-intervention-credit: LLM-serving intervention credit simulation results.
• consistency-aware-judgments: consistency-aware ranking/judgment data.
• frontier-allocation-metrics: budgeted inference/frontier allocation metrics.
• mwfas-heuristic-metrics: minimum-weighted feedback arc set heuristic metrics.
• ranking-fas-results: ranking by feedback arc set results.

This release concerns natural-language optimization problem instantiation and retrieval/grounding readiness. It does not supersede any existing dataset.

Ownership, Provenance, and License

Soroush Vahidi is the dataset author for this derived metrics release.

The released derived metrics, annotations, and release documentation are licensed under CC BY-NC-SA 4.0. This license applies solely to the output structural metrics and metadata contained within this repository; it does not relicense NLP4LP or OptMATH. Upstream sources remain subject to their own access controls and licenses.

Observed upstream NLP4LP metadata at build time:

• Canonical source: https://huggingface.co/datasets/udell-lab/NLP4LP
• Maintainer: Udell Lab
• Hugging Face license tag: cc-by-nc-sa-4.0
• Access: gated

Observed upstream OptMATH metadata at build time:

• Canonical source: https://huggingface.co/datasets/shushulei/OptMATH-Train (and Github: shushulei/OptMATH)
• Maintainer: Shushulei / OptMATH Contributors
• Hugging Face license tag: unknown (Github repository code licensed under Apache-2.0)
• Dataset License Status: DATASET_LICENSE_UNCLEAR_BUT_DERIVED_METRICS_SAFE (As the OptMATH-derived configs in this release only contain abstract validation counts and evaluation recalls without redistributing problem descriptions, formulations, or array values, the derived metrics release falls securely within fair use boundaries without conflicting with upstream restrictive or unstated terms).

Associated Paper

Direct associated paper:

Soroush Vahidi. Retrieval-Assisted Instantiation of Natural-Language Optimization Problems. Preprint records: SSRN DOI 10.2139/ssrn.6434886; Research Square DOI 10.21203/rs.3.rs-10463941/v1.

Note on v1.1: The optmath_external_validation configuration is a later external-validation extension of the dataset supporting the same research pipeline; it should not be interpreted as a result necessarily reported in the originally published associated preprint.

Citation Policy

Cite this dataset when using the released rows or metrics. Cite the associated paper when using or discussing the methodology or scientific results. Cite both when using the released data and relying materially on the associated methodology/results. Cite NLP4LP/OptiMUS for upstream benchmark provenance.

Dataset citation:

@dataset{vahidi_nl_optimization_instantiation_metrics_2026,
  title = {Natural-Language Optimization Instantiation Metrics},
  author = {Vahidi, Soroush},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SoroushVahidi/nl-optimization-instantiation-metrics}},
  note = {Derived metrics release; no dataset DOI assigned}
}

Associated paper citation:

@misc{vahidi_retrieval_assisted_instantiation_2026,
  title = {Retrieval-Assisted Instantiation of Natural-Language Optimization Problems},
  author = {Vahidi, Soroush},
  year = {2026},
  doi = {10.2139/ssrn.6434886},
  note = {Preprint; Research Square DOI: 10.21203/rs.3.rs-10463941/v1}
}

Upstream benchmark citation:

@inproceedings{ahmaditeshnizi2024optimus,
  title = {{O}pti{MUS}: Scalable Optimization Modeling with ({MI}){LP} Solvers and Large Language Models},
  author = {Ahmaditeshnizi, Ali and Gao, Wenzhi and Udell, Madeleine},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning},
  pages = {577--596},
  year = {2024},
  volume = {235},
  series = {Proceedings of Machine Learning Research},
  publisher = {PMLR},
  url = {https://proceedings.mlr.press/v235/ahmaditeshnizi24a.html}
}