SoroushVahidi/frontier-allocation-metrics

Frontier Allocation Metrics: Per-Query Cost, Latency, and Accuracy Outcomes for Budgeted Large-Language-Model (LLM) Inference

This is a metrics-only research dataset for budgeted LLM (Large Language Model) inference allocation experiments — that is, experiments that study how a fixed "budget" of inference calls should be spent across a search/retry process to trade off cost, latency, and accuracy. It contains sanitized per-query outcome measurements from matched-budget runs over two provider/model snapshots and four public reasoning benchmarks. It does not contain benchmark question text, answer text, prompts, raw model completions, API payloads, private endpoints, account metadata, or provider request identifiers.

Provenance in one line: derived evaluation metrics, generated by Soroush Vahidi, computed over publicly available third-party benchmarks (GSM8K, MATH-500, GPQA-Diamond, StrategyQA) and commercial provider outputs (Azure OpenAI, Google Vertex Gemini). The benchmark problem text and the raw provider completions themselves are not redistributed — see "Exclusions" below.

Current version: v1.1 (published 2026-08-16, Hugging Face revision b6367eb0…). v1.1 is a strictly additive update over v1 — see "Version History".

The planned canonical Hugging Face repository is SoroushVahidi/frontier-allocation-metrics.

What One Row Means

In per_query_outcomes, one row represents one evaluated method on one opaque benchmark example under one provider/model and a fixed budget of six logical inference calls. A row is not a prompt, response, or benchmark example; it is a compact outcome record with method, provider/model, benchmark, opaque example ID, token counts, latency, estimated cost, correctness, and validation metadata. The three other configs summarize, aggregate, or trace this same underlying run set at different levels of detail (see "Dataset Structure" below).

Dataset Size

Config	Rows	Description
per_query_outcomes	7,184	Primary method-level outcome table (one row per evaluated method × example × provider/model).
aggregate_matrix	32	Provider × model × benchmark × method summaries, recomputed directly from per_query_outcomes.
search_trace_summary	1,992	Per-run execution summary for Google Vertex Gemini runs (added in v1.1).
search_trace_nodes	3,092	Node-level, action-by-action search-tree traces (added in v1.1).

Total: 12,300 rows across 4 configs.

What Is Included

• Providers/model snapshots: Azure OpenAI gpt-4.1-mini; Google Vertex Gemini gemini-2.5-flash.
• Benchmarks: GSM8K, MATH-500, GPQA-Diamond, StrategyQA.
• Methods: Frontier (direct_reserve_semantic_frontier_v2), L1 (external_l1_max), S1 (external_s1_budget_forcing), and TALE (external_tale_prompt_budgeting). The internal audit also tracked a Failure-Trace Allocator (FTA) method, but the public aggregate_matrix table is intentionally recomputed only from rows actually present in per_query_outcomes.
• Sanitized per-query and per-search-node measurements: token counts, latency, estimated cost, correctness, and search-tree topology/priority signals — see the data dictionaries below.

What Is Not Included

• Benchmark question text, answer text, or gold answers (for any of the four benchmarks).
• Raw model prompts or completions.
• API payloads, private endpoints, or provider account metadata.
• Cohere and Fireworks provider rows — excluded from v1/v1.1 under a reduced scope adopted after provider-terms review. Fireworks × GPQA-Diamond additionally had protocol nonconvergence in the internal audit and is not represented as performance data at all.
• Original GPQA-Diamond identifiers, question text, or answer choices — GPQA-Diamond rows use only opaque, project-local example identifiers.

Dataset Structure / Schema

per_query_outcomes (7,184 rows) and aggregate_matrix (32 rows)

Column	Type	Meaning
benchmark	string	Public benchmark name.
example_uid	string	Stable, release-local opaque example identifier generated by a private-key HMAC. Raw IDs and the mapping key are not distributed.
provider	string	Public provider label.
model	string	Public model/snapshot identifier.
method	string	Public method label: Frontier, L1, S1, or TALE.
method_source_id	string	Original implementation identifier, for reproducibility.
budget_logical_calls	int64	Logical inference-call budget. All v1/v1.1 rows use 6 — not necessarily six raw HTTP requests.
input_tokens / output_tokens / total_tokens	int64	Logged/estimated token counts.
latency_seconds	float64	Logged method-level wall-clock latency, not a provider SLA figure.
estimated_cost_usd	float64	Estimated USD cost from logged token usage and pricing assumptions — not an audited invoice.
exact_match	bool	Benchmark-normalized correctness indicator; gold answers are not released.
gold_in_tree	bool	Whether the gold answer appeared anywhere in the explored candidate set; the answer itself is not released.
parse_extraction_failure	bool	Whether answer extraction/parsing from the model output failed.
trust_tier / validation_status	string	Internal audit trust tier and validation status.
acquisition_date	string	Coarse YYYY-MM-DD run date.
source_cell_id	string	Public-safe provider/model/benchmark cell identifier (no filesystem paths).

aggregate_matrix has the same conceptual columns pre-aggregated to counts/rates/means per provider × model × benchmark × method cell (e.g. exact_match_rate, mean_latency_seconds, total_estimated_cost_usd) — see the live dataset card's full per-column listing for the exact aggregate schema.

search_trace_summary (1,992 rows) — added in v1.1

One row per executed run (joins to per_query_outcomes/search_trace_nodes via example_uid / trace_uid), restricted to Google Vertex Gemini on MATH-500 and GPQA-Diamond. Adds trace_uid (unique run identifier), has_trace (whether a search tree was actually expanded — true for Frontier and L1), and repeats the core outcome columns above for that run.

search_trace_nodes (3,092 rows) — added in v1.1

Node-level, action-by-action traces of the search tree for Frontier and L1 methods. Key columns: trace_uid (links to search_trace_summary), node_id/parent_node_id (anonymized tree identifiers), depth, action_type (expand or direct_reserve), and a set of float-valued search metrics (priority, continuation_value, diversity_bonus, duplicate_cost, coverage_gain, semantic_overlap, plausibility_score, base_priority_score), plus policy flags (force_explore, gate_intervened) and expansion_order. Fully text-free.

(Full per-column tables for all four configs, exactly as published, are in the live dataset card — reproduced above are the columns most researchers will use first.)

Quickstart

from datasets import load_dataset

# Primary per-query outcome table
outcomes = load_dataset("SoroushVahidi/frontier-allocation-metrics", "per_query_outcomes")
print(outcomes)

# Pre-aggregated provider x model x benchmark x method summary
aggregate = load_dataset("SoroushVahidi/frontier-allocation-metrics", "aggregate_matrix")
print(aggregate)

# v1.1 additions: search-tree traces (Google Vertex Gemini runs only)
trace_summary = load_dataset("SoroushVahidi/frontier-allocation-metrics", "search_trace_summary")
trace_nodes = load_dataset("SoroushVahidi/frontier-allocation-metrics", "search_trace_nodes")

Config names verified against the live README's configs: YAML block and the local publication record's config list; row counts verified against row_counts.json and the v1.1 publication record.

Research Use Cases

• Reproducing the cost/accuracy/latency comparisons in the associated paper without rerunning paid provider API calls.
• Studying cost, latency, and accuracy tradeoffs under a fixed inference-call budget, across methods and benchmarks.
• Per-query method-selection / routing research (e.g., when does one budgeted-inference method outperform another, and under what token/latency profile).
• Search-tree topology and stopping-behavior analysis using search_trace_summary / search_trace_nodes (branching factor, depth, priority-score distributions, gate_intervened triggers).

Not intended as a general model-quality leaderboard — see "Limitations".

Generation Methodology

Derived by running four budgeted-inference methods against four public reasoning benchmarks under two commercial provider/model snapshots, at a fixed logical-call budget of 6, then sanitizing and aggregating the resulting outcome logs (see the source repository's experiments/ and paper_performance_evaluation/ for the full protocol). Example identifiers are generated via a private-key HMAC so that individual benchmark items cannot be recovered from the release.

Provenance and Ownership

• v1 (2 configs: per_query_outcomes, aggregate_matrix): generated by Soroush Vahidi from the frontier-allocation-for-budgeted-llm-inference source repository.
• v1.1 additions (search_trace_summary, search_trace_nodes): also generated by Soroush Vahidi, over the same underlying runs, adding search-tree-level detail. The exact source-code commit that produced these two new configs could not be traced in the local frontier-allocation-for-budgeted-llm-inference git repository as of this audit — the repo's checked-out history predates the v1.1 publication date, and the object is not present in that repo or any other local repository checked. This is a documentation/traceability gap, not a data problem: the v1.1 Parquet files themselves are hash-verified against the published Hugging Face revision and passed the project's own security and source-integrity checks (security_result: PASS, source_integrity_result: PASS, gpqa_text_free_verification: PASS, from the local publication record). Until the exact commit is recovered, treat the code provenance of the v1.1 additions as unconfirmed while treating the data itself as verified-current.
• Third-party inputs: GSM8K, MATH-500, GPQA-Diamond, StrategyQA (public benchmarks, not redistributed); Azure OpenAI and Google Vertex Gemini (commercial provider outputs, sanitized into metrics only).

Relationship to Related Datasets

• SoroushVahidi/lafc-evict — cache-eviction candidate supervision; unrelated research problem.
• SoroushVahidi/module-intervention-credit — LLM-serving scheduler module-intervention data; different task and schema.
• SoroushVahidi/consistency-aware-judgments — pairwise LLM judgments for retrieval consistency; different representation and research question.
• SoroushVahidi/scidocs — third-party BEIR/SciDocs mirror; no content overlap.
• SoroushVahidi/lafc-evict-sample — synthetic workflow test artifact; unrelated.

Limitations

• Point-in-time measurements from two specific provider/model snapshots and one fixed logical-call budget (6) — do not use as a general model-quality leaderboard.
• estimated_cost_usd is a pipeline estimate from logged token usage, not an audited invoice.
• latency_seconds may include client-side/runtime overhead, not a provider SLA figure.
• Opaque example IDs are not benchmark IDs and should not be reverse-engineered.
• The exact source commit for the v1.1 search_trace_* additions is not currently traceable locally (see "Provenance and Ownership" above) — a documentation gap the maintainer intends to close.

Version History

• v1: per_query_outcomes, aggregate_matrix (2 configs).
• v1.1 (2026-08-16, current): adds search_trace_summary, search_trace_nodes (2 new configs), strictly additive — no existing v1 values were altered (confirmed via matching hashes for the two original Parquet files between the v1 and v1.1 publication records).

Related Resources

• Associated paper: Soroush Vahidi. Selective Deferral for Budgeted LLM Answer Selection: Failure-Trace Signals under Matched-Budget Evaluation. Research Square preprint, 2026. DOI: 10.21203/rs.3.rs-9783817/v1. URL: https://www.researchsquare.com/article/rs-9783817/latest
• Source code / project repository: https://github.com/SoroushVahidi/frontier-allocation-for-budgeted-llm-inference

Citation

Cite the dataset when using the released data; cite the paper when discussing the methodology or reported results; cite both when your work uses the data and relies on the methodology.

Dataset:

@dataset{vahidi2026_frontier_allocation_metrics,
  author    = {Vahidi, Soroush},
  title     = {Frontier Allocation Metrics: Per-Query Cost, Latency, and Accuracy Outcomes for Budgeted LLM Inference},
  year      = {2026},
  version   = {v1.1},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/SoroushVahidi/frontier-allocation-metrics},
  license   = {CC-BY-4.0}
}

Paper:

@article{vahidi2026_selective_deferral,
  author = {Vahidi, Soroush},
  title  = {Selective Deferral for Budgeted LLM Answer Selection: Failure-Trace Signals under Matched-Budget Evaluation},
  year   = {2026},
  journal = {Research Square (preprint)},
  doi    = {10.21203/rs.3.rs-9783817/v1}
}

No dataset-specific DOI was found — only the paper's Research Square preprint DOI above is verified. If a dataset DOI is minted later (e.g. via a Zenodo/DataCite archive), add it here.

License

CC BY 4.0 for the released derived metrics. This license does not relicense provider APIs/models, prompts, raw model outputs, or the underlying benchmark text — follow the upstream terms for GSM8K, MATH-500, GPQA-Diamond, and StrategyQA if you combine this dataset with upstream benchmark material. Code in the associated source repository is separately licensed under the MIT License (see its CITATION.cff/LICENSE).