SoroushVahidi/frontier-allocation-metrics

Frontier Allocation Metrics: Per-Query Cost, Latency, and Accuracy Outcomes for Budgeted LLM Inference

This is a metrics-only research dataset for budgeted LLM inference allocation experiments. It contains sanitized per-query outcome measurements from matched-budget runs over two provider/model snapshots and four 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.

The planned canonical Hugging Face repository is SoroushVahidi/frontier-allocation-metrics. This local staging release is version v1.

Immutable v1 scientific release revision:

8f2093cf646fe61064321e6464d852cc55ebe0b4

For version-specific citation, cite the repository together with version v1 and this Hugging Face revision. Later documentation-only commits may update the dataset card, but the Parquet data files for v1 are identified by the scientific release revision above.

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.

Configurations

• per_query_outcomes: 7,184 rows. Primary method-level outcome table.
• aggregate_matrix: 32 rows. Provider × model × benchmark × method summaries recomputed directly from per_query_outcomes.
• search_trace_summary: 1,992 rows. Detailed per-query execution summary for Google Vertex Gemini runs, recording general metadata and whether a search tree was expanded.
• search_trace_nodes: 3,092 rows. Node-level action-by-action search tree traces, logging step depth, priority scores, and strategic decisions. Fully text-free and linkable to summary tables via unique trace hashes.

Included Scope

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 conclusive internal audit also tracked Failure-Trace Allocator (FTA) aggregate behavior, but this v1 public aggregate table is intentionally recomputed only from rows present in per_query_outcomes.

Exclusions

Cohere and Fireworks rows are excluded from v1 under the reduced scope approved after provider-terms review. Fireworks × GPQA-Diamond also had protocol nonconvergence in the conclusive audit and is not represented as performance data.

GPQA-Diamond is included only as derived metrics with opaque project-local example identifiers. Original GPQA IDs, question text, answer choices, and answer labels are not distributed.

Data Dictionary

Column	Type	Role	Meaning
benchmark	string	metadata	Public benchmark name.
example_uid	string	opaque ID	Stable release-local opaque example identifier generated by private-key HMAC; raw IDs and mapping key are not distributed.
provider	string	metadata	Public provider label.
model	string	metadata	Public model/snapshot identifier recorded by the run.
method	string	method	Public method label: Frontier, L1, S1, or TALE.
method_source_id	string	method provenance	Original implementation identifier for reproducibility.
budget_logical_calls	int64	condition	Logical inference-call budget. All v1 rows use 6. This is not necessarily six raw HTTP requests.
input_tokens	int64	measurement	Logged/estimated input-token count.
output_tokens	int64	measurement	Logged/estimated output-token count.
total_tokens	int64	measurement	Logged total token count.
latency_seconds	float64	measurement	Logged method-level latency/wall-clock seconds, not a provider SLA.
estimated_cost_usd	float64	measurement	Estimated USD API cost from logged token usage and pricing assumptions, not audited invoice cost.
exact_match	bool	outcome	Benchmark-normalized correctness indicator; gold answers are not released.
gold_in_tree	bool	diagnostic	Whether the gold answer was present among the explored candidate set/tree; the answer itself is not released.
parse_extraction_failure	bool	diagnostic	Whether answer extraction/parsing from model output failed.
trust_tier	string	audit metadata	Canonical audit trust tier.
validation_status	string	audit metadata	Canonical validation status. COMPLETE_VALIDATED_OFFLINE_FTA records the Azure OpenAI × GPQA-Diamond offline-FTA validation caveat.
acquisition_date	string	metadata	Coarse YYYY-MM-DD date inferred from canonical run identifiers.
source_cell_id	string	provenance	Public-safe provider/model/benchmark cell identifier; no filesystem paths.

Data Dictionary: search_trace_summary

Column	Type	Role	Meaning
trace_uid	string	unique ID	Unique execution run SHA256 identifier mapping 1-to-1 with a specific model/method run under a budget of 6.
benchmark	string	metadata	Public benchmark name: MATH-500 or GPQA-Diamond.
example_uid	string	opaque ID	Stable release-local opaque example identifier mapping to raw IDs; joins directly with per_query_outcomes and search_trace_nodes.
provider	string	metadata	Public provider label (Google Vertex Gemini).
model	string	metadata	Public model label (gemini-2.5-flash).
method	string	method	Clean publication method name: Frontier, L1, S1, or TALE.
method_source_id	string	provenance	Raw method source code ID.
budget_logical_calls	int64	condition	Logical inference budget. All rows use 6.
input_tokens	int64	measurement	Logged input token count for this run.
output_tokens	int64	measurement	Logged output token count for this run.
total_tokens	int64	measurement	Logged total token count for this run.
latency_seconds	float64	measurement	Wall-clock seconds.
estimated_cost_usd	float64	measurement	Cost estimate in USD.
exact_match	bool	outcome	Correctness status.
gold_in_tree	bool	diagnostic	Whether the correct answer was generated somewhere in the tree search.
parse_extraction_failure	bool	diagnostic	Parsing error indicator.
trust_tier	string	metadata	Trust rating (B).
validation_status	string	metadata	Validation status (COMPLETE_VALIDATED).
acquisition_date	string	metadata	Run date (2026-07-05).
source_cell_id	string	provenance	Public-safe cell name.
has_trace	bool	flag	True if the search tree was actually expanded (Frontier and L1 methods).

Data Dictionary: search_trace_nodes

Column	Type	Role	Meaning
trace_uid	string	link ID	Unique execution run hash; joins directly with search_trace_summary.
example_uid	string	opaque ID	Opaque example ID for direct independent joins.
benchmark	string	metadata	Benchmark label.
method	string	method	Method name (Frontier or L1).
budget_logical_calls	int64	condition	Budget size (6).
node_id	string	node ID	Anonymized node/branch identifier (e.g. div_0).
parent_node_id	string	node ID	Anonymized parent node identifier.
depth	int64	search topology	Depth in search tree/branching path.
action_type	string	search policy	Decision made: expand or direct_reserve.
priority	float64	search metrics	Priority value calculated for this branch during tree exploration.
continuation_value	float64	search metrics	Continuation model score.
diversity_bonus	float64	search metrics	Novelty bonus calculated.
duplicate_cost	float64	search metrics	Duplicate penalty applied.
coverage_gain	float64	search metrics	Predicted coverage increase.
semantic_overlap	float64	search metrics	Semantic overlap deduction.
force_explore	bool	policy flag	Forced exploration override flag.
gate_intervened	bool	policy flag	Decisive meta-level gate override indicator.
plausibility_score	float64	search metrics	Model node plausibility probability.
base_priority_score	float64	search metrics	Baseline node priority before penalties.
expansion_order	int64	search topology	Chronological expansion index of this node in the exploration process.
strategy_family	string	search provenance	Search algorithm category applied.

Budget, Cost, Latency, and Correctness Caveats

The logical-call budget is fixed at 6 across v1. Methods can use the project protocol's logical budget differently, so budget_logical_calls=6 should not be read as six identical model invocations. estimated_cost_usd is a pipeline estimate, not an audited billing record. latency_seconds is a logged method-level runtime measurement and may include client-side/runtime overhead. exact_match is benchmark-normalized and benchmark-family-specific; raw questions, answers, and model outputs are intentionally excluded.

Scientific Use

This dataset supports research on budgeted LLM inference allocation, cost/accuracy/latency tradeoffs, per-query method selection, historical provider/model behavior, and reproducibility of Frontier Allocation method comparisons without rerunning paid API experiments. It saves future researchers thousands of paid model evaluations while avoiding redistribution of benchmark content or model outputs.

The structural trace configurations (search_trace_summary and search_trace_nodes) enable deeply specific new research directions, including:

• Search Tree Topology Analysis: Studying how branching factors, exploration depth, and node counts vary dynamically across MATH and GPQA problem domains.
• Budget Allocation Efficiency: Analyzing the precise cost/latency/accuracy tradeoffs of incremental node expansion and the dynamic execution cost of priority-guided search strategies.
• Stopping and Deferral Behavior: Assessing the mechanistic triggers (such as gate_intervened or force_explore) that determine optimal search termination under strict resource caps.
• Priority-Guided Evaluation: Probing the actual math-problem solving and GPQA-solving trajectory choices, investigating the distribution of priority and plausibility scores over successful versus failing exploration pathways.
• Reproducibility of Structural Frontier Behavior: Providing complete, exact, reproducible chronological search paths (expansion_order) without exposing restricted benchmark or model-generated text.

Non-Intended Uses and Limitations

Do not use this as a universal model-quality leaderboard. The measurements are point-in-time outcomes from specific provider/model snapshots, benchmark subsets, method implementations, and one logical-call budget. Do not use opaque IDs as benchmark IDs or attempt to reverse them. Do not treat model-generated outcomes as human annotations.

How This Dataset Differs from Existing Datasets

This release differs from raw benchmarks because it does not redistribute question or answer text. It differs from aggregate leaderboards and pricing tables because it provides per-query cost, token, latency, and correctness measurements under matched-budget method comparisons. It differs from RouterBench, LLMRouterBench, RouteLLM, and R2-Router/R2-Bench-style resources by focusing on compact metrics-only Frontier Allocation experiment outcomes over selected reasoning benchmarks and provider snapshots, rather than broad model-routing corpora, preference-router training data, or output-length-budget routing datasets.

No claim is made that this dataset is first, largest, unique, comprehensive, or state of the art.

Relation to Other SoroushVahidi Hugging Face Datasets

• SoroushVahidi/lafc-evict: cache-eviction candidate supervision; unrelated scientific problem.
• SoroushVahidi/module-intervention-credit: LLM-serving scheduler module intervention/credit data; different task and schema.
• SoroushVahidi/consistency-aware-judgments: pairwise LLM judgments for information-retrieval consistency analysis; different representation and research question.
• SoroushVahidi/scidocs: third-party BEIR/SciDocs mirror; this dataset does not redistribute SciDocs content.
• SoroushVahidi/lafc-evict-sample: synthetic workflow artifact; unrelated.

Associated Paper

Direct 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

The paper describes the matched-budget selective-deferral methodology and reports related Frontier/FTA analyses. 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 associated methodology/results.

License and Attribution

Released derived metrics are provided under CC BY 4.0 to the extent controlled by this project. This license does not relicense provider APIs/models, prompts, raw model outputs, or underlying benchmark text. Users should follow upstream benchmark terms for GSM8K, MATH-500, GPQA-Diamond, and StrategyQA when resolving opaque IDs or combining this dataset with upstream benchmark material.

Citation

Vahidi, S. (2026). Frontier Allocation Metrics: Per-Query Cost, Latency, and Accuracy Outcomes for Budgeted LLM Inference (v1). Hugging Face dataset. https://huggingface.co/datasets/SoroushVahidi/frontier-allocation-metrics

@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},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/SoroushVahidi/frontier-allocation-metrics},
  license = {CC-BY-4.0}
}

Source code/project repository: https://github.com/SoroushVahidi/frontier-allocation-for-budgeted-llm-inference