SoroushVahidi/llm-serving-selector-regret

LLM-Serving Selector Regret

LLM-Serving Selector Regret is a metrics-only research dataset for studying learned policy selection in LLM-serving schedulers. It contains derived selector/oracle/regret objects generated by Soroush Vahidi's research workflow, not raw request traces.

Creator / Provider

Dataset creator/provider: Soroush Vahidi.

The released selector/regret and policy-suitability metrics were generated by Soroush Vahidi's research workflow. Underlying third-party workload traces remain subject to their original providers and licenses; this dataset does not claim ownership of those upstream traces.

Dataset Structure

This standalone dataset has five configs:

Config	Rows	Columns	Row meaning
selector_regret	3,852	160	One workload/window with context features, 27 candidate-policy reward values, oracle fields, split metadata, and regret-related decision fields.
policy_suitability	95,364	10	One (model_id, split, window_id, policy_name) record for the validation-selected finalist selector.
selector_v2_registry	640	51	One (window, policy) pair with 48 numerical simulation outcome metrics, 49 window features, and split metadata — training-oriented selector data for the 8-policy Option B scope.
selector_v3_windows	2,568	98	One retained window with domain/scenario metadata, a 7-way multidomain train/OOD split, 82 engineered features (49 shared with selector_regret plus a new 33-column feat_v3_* rolling-window block), and 3 causal-discriminability labels.
selector_v3_policy_vectors	20,544	52	One (window_idx, policy) pair (8 policies per window) with simulation outcome metrics, using the same 8-policy library and metric schema as selector_v2_registry, over a disjoint, ~32x larger, 3-domain window set.

This dataset is intentionally separate from SoroushVahidi/llm-serving-scheduler-baselines. That dataset contains fixed scheduler outcome rows. This dataset contains selector/oracle/regret and suitability objects that require additional intermediate features and learned-selector outputs and are not reconstructible from the scheduler-baselines release alone.

Quickstart

from datasets import load_dataset

# Wide-format policy benchmark (27 policies, oracle/regret labels)
selector_regret = load_dataset("SoroushVahidi/llm-serving-selector-regret", "selector_regret")

# Long-format finalist-selector suitability scoring
policy_suitability = load_dataset("SoroushVahidi/llm-serving-selector-regret", "policy_suitability")

# v2 calibration pilot (8-policy Option B scope, long format)
selector_v2_registry = load_dataset("SoroushVahidi/llm-serving-selector-regret", "selector_v2_registry")

# v3 multi-domain window features + causal-discriminability labels (join key: window_idx)
selector_v3_windows = load_dataset("SoroushVahidi/llm-serving-selector-regret", "selector_v3_windows")

# v3 per-(window, policy) outcome metrics (join key: window_idx, 8 rows per window)
selector_v3_policy_vectors = load_dataset("SoroushVahidi/llm-serving-selector-regret", "selector_v3_policy_vectors")

print(selector_regret)

Which config should I use? If you're training or evaluating a policy selector, start with selector_v2_registry (smaller, well-established) or selector_v3_windows + selector_v3_policy_vectors (larger, multi-domain, joined on window_idx) — see their sections below for the relationship between v2 and v3. If you're benchmarking a fixed set of 27 policies against an oracle, use selector_regret. If you're evaluating how well a specific trained selector model performs, use policy_suitability.

selector_regret Semantics

A selector_regret row represents one workload/window in a 27-policy full-information selector benchmark. Column groups are:

• workload/window identifiers: window_id, source fields, temporal fields, and split fields;
• context and workload features: columns prefixed with feat_ plus workload parameter columns;
• policy reward vector: one numeric reward column for each candidate policy, including fifo, edf, weighted_shortest_processing, scorpio_style_slo_guard, kv_constrained_online, and the rest of the 27-policy V2 library;
• oracle decision fields: oracle_v2_anwg, best_policy, second_best_margin, and meaningful_margin;
• V1 comparison fields: oracle_v1_anwg and v2_minus_v1_gain;
• epsilon-optimality fields: epsilon_optimal_count_0.001, epsilon_optimal_count_0.005, and epsilon_optimal_count_0.01.

The objective is arrival_normalized_weighted_goodput; higher is better. For a window w and policy p, with reward R(w,p), the V2 oracle reward is:

oracle_v2_anwg(w) = max_p R(w,p).

The per-policy true regret used by the suitability table is absolute, non-normalized reward loss:

true_regret(w,p) = oracle_v2_anwg(w) - R(w,p).

This value is non-negative up to numerical tolerance. Ties are handled by the underlying tabular maximum/argmax behavior used during generation; epsilon-optimality counts expose near-tie structure at thresholds 0.001, 0.005, and 0.01. The release validation confirms every benchmark window has all 27 policy reward values available.

policy_suitability Semantics

A policy_suitability row represents one candidate policy evaluated for one non-TRAIN workload/window by the validation-selected finalist selector model.

Columns:

• model_id: selected finalist selector model identifier.
• split: evaluation split.
• window_id: workload/window key that joins to selector_regret.window_id.
• policy_name: candidate policy being scored.
• ranking_score: model score used to rank policies for the window.
• probability_epsilon_optimal_proxy: softmax-style proxy over policy scores for epsilon-optimality inspection; it is a diagnostic score, not a calibrated probability guarantee.
• predicted_regret: non-negative model-implied regret proxy relative to the highest model score in that window.
• true_reward: realized arrival_normalized_weighted_goodput for this policy/window.
• true_regret: oracle_v2_anwg - true_reward from the selector-regret table.
• selected: 1 if the finalist selector selected this policy for the window, else 0.

policy_suitability intentionally excludes TRAIN windows. It covers 3,532 non-TRAIN windows x 27 policies = 95,364 rows.

selector_v2_registry Semantics

selector_v2_registry is the selector-v2 calibration pilot — a curated, training-oriented subset derived from the same simulation infrastructure and trace families as selector_regret, but structured differently for training a learned policy selector model.

Structural relationship to selector_regret: Both share the primary objective (arrival_normalized_weighted_goodput) and the same upstream workload families (BurstGPT, AzureLLM 2023). However:

• selector_regret uses a wide format (1 row per window, 27 policy reward columns) for benchmarking.
• selector_v2_registry uses a long format (1 row per window×policy pair) for training.

They are related but non-redundant. Registry rows are not reconstructible from selector_regret (different window sets, different policy subsets), and selector_regret rows are not reconstructible from the registry.

selector_v2_registry row definition

A selector_v2_registry row represents one simulation outcome when a single scheduling policy is applied to a single workload window.

• Window: A contiguous time slice of a workload trace (synthetic or real-trace), defined by an arrival range and row index in the source JSONL.
• Policy: One of 8 non-faithful scheduling policies (Option B scope).
• Outcome: 48 numerical metrics computed by the simulation engine.

Columns

Identifiers

• window_idx — Sequential window identifier (0-based)
• group_key — Grouping key for split-atomicity
• split — TRAIN / VALIDATION / ID_TEST / OOD_TEST
• split_group_key — Split-group identifier for leakage-safe grouping
• policy_name — Scheduling policy evaluated

Window features (49 feat_ columns)* Context features extracted from each window: arrival rates, burstiness coefficients, queue lengths, priority distributions, resource configuration (GPU count, KV capacity, sequence capacity, token budget), saturation load estimates, tightness fractions, slack metrics, and disaggregated/multi-instance topology features where applicable.

Simulation outcome metrics (48 metric_ columns)* Numerical outcomes from policy evaluation: arrival_normalized_weighted_goodput, latencies (mean, median, p50, p95, p99), throughput (request, token, SLO success), SLO attainment/violation rates, queue lengths, GPU utilization, admission/rejection rates, and event counts.

Some metrics are null for certain workload types:

• Disaggregated topology metrics (prefill/decode queue utilization, bridge queue metrics) are null for monolithic workloads.
• Some latency/throughput/SLO metrics are null for a small subset of synthetic workloads (up to 4.1% of rows).

Labels / targets

• primary_objective_classification — Discernibility classification per window: STRONGLY_DISCRIMINATIVE, NEAR_TIE, or ALL_COMPLETE_OR_EFFECTIVELY_TIED
• primary_objective_best_policy — The best-performing policy of the 8 for this window (by ANWG)
• primary_objective_max_min_spread — ANWG gap between best and worst of the 8 policies

Policy set (Option B scope)

1. fifo
2. edf
3. scorpio_style_slo_guard
4. admission_control
5. weighted_shortest_processing
6. estimated_service_time_first
7. best_fit
8. multi_bin_batching

The following are explicitly excluded from this config: vllm_faithful, vllm_chunked_prefill_faithful, sarathi_faithful, distserve_faithful, tetriinfer_paper_reimplementation, llumnix_faithful.

Calibration

SLO multiplier = 2.0, policy-independent reference model. Trace-level OOD reservation: last 15% of rows (by arrival time) in each source trace are reserved exclusively for OOD_TEST, disjoint from historical pool by row-index construction.

Intended use

This config is intended for training learned selector models to choose which of the 8 policies to apply per workload window, given window-level features. It complements selector_regret (which benchmarks selector models across 30 policies) and policy_suitability (which evaluates a finalist selector against non-TRAIN windows).

Split distribution

Split	Windows	Policies/Window	Rows
TRAIN	31	8	248
VALIDATION	18	8	144
ID_TEST	13	8	104
OOD_TEST	18	8	144
Total	80	8	640

Known issue: 19 cross-split row-range overlap pairs across 27 distinct windows in TRAIN/VALIDATION/ID_TEST pools. OOD_TEST is structurally clean (reserved by row index).

Provenance

Generated from llm-serving-heuristic-evolution at commit b698e49df7954ad6fb8e3026d27b091e971248bd ("Reconcile Selector v2 split-leakage fix with integration branch's c8aee12").

Source experiment: selector_v2_calibrated_pilot_20260720T163235Z. Validation manifests passed for dataset construction, quality gates (7/7), and leakage audit.

selector_v3_windows and selector_v3_policy_vectors Semantics

These two configs are a same-pipeline, multi-domain scale-up of selector_v2_registry, generated one day later by the same experimental line (selector_v2_overnight_20260720T235405 → selector_v2_ood_conclusive_20260721T133408Z → this release). They study whether a learned policy-selection model generalizes across LLM-serving trace domains (Azure-2023, BurstGPT, and synthetic stress scenarios), not just within one domain.

Relationship to selector_v2_registry — read this before using both

selector_v3_policy_vectors uses the identical metric schema and 8-policy library as selector_v2_registry (same generating pipeline, commit c8aee129f553f8dc3ede99eac60d5b14484beb41), at roughly 32x the window count (2,568 vs. 640) and across 3 explicit source domains instead of a narrower scope.

This is not an independent metric design, and it is not a replacement. The two configs have zero overlapping windows (disjoint naming schemes; 0/20,544 exact row matches against the live selector_v2_registry config, confirmed by direct join) and different split granularity (7-way vs. v2's 4-way). selector_v2_registry is not superseded and remains published unchanged — a reader interested in the smaller, earlier v2 window set still needs it; a reader interested in cross-domain robustness needs v3. Do not merge the two into one versioned table.

Label-vocabulary note: selector_v3_windows.primary_objective_classification uses the three values STRONGLY_DISCRIMINATIVE, MODERATELY_DISCRIMINATIVE, NEAR_TIE — the middle category name differs from selector_v2_registry.primary_objective_classification's ALL_COMPLETE_OR_EFFECTIVELY_TIED. Do not assume the two columns share an identical three-way taxonomy across configs.

selector_v3_windows row definition (2,568 rows, 98 columns)

One row per retained simulated scheduling window. window_idx is the join key to selector_v3_policy_vectors (1:1).

• Domain/scenario metadata: source_trace (azure_llm_2023, burstgpt, or synthetic), dataset_family (real_trace or controlled_stress), a named scenario shape (5 real-trace shapes plus 5 named synthetic stress scenarios: closely-spaced/same-arrival heterogeneous clusters, KV-pressure admission ordering, long-prefill overlap, admission-reorder boundary).
• Split: a 7-way taxonomy — TRAIN, VALIDATION, ROBUST_DEV, ID_TEST, CROSS_SOURCE_OOD, TEMPORAL_OOD, FINAL_OOD — materially finer than selector_v2_registry's 4-way scheme, separating "OOD" into cross-source, temporal, and held-out-final variants.
• Engineered features (82 columns): 49 columns reuse the same feat_* vocabulary already published in selector_regret; the remaining 33 columns are a new, fully-populated feat_v3_* block (rolling 1s/5s/20s/60s arrival/work rates, recent work/slack percentiles, negative-laxity fractions, estimated KV pressure, queue-growth-rate) with no analogue elsewhere in this dataset.
• Causal-discriminability labels: primary_objective_classification (see label-vocabulary note above), primary_objective_best_policy (best of the 8 policies for this window by ANWG), primary_objective_max_min_spread (ANWG gap between best and worst of the 8 policies).

Split distribution: TRAIN 968, VALIDATION 160, ROBUST_DEV 160, ID_TEST 312, CROSS_SOURCE_OOD 788, TEMPORAL_OOD 152, FINAL_OOD 28 (sums to 2,568). Domain distribution: azure_llm_2023 1,168 windows, burstgpt 760, synthetic 640.

selector_v3_policy_vectors row definition (20,544 rows, 52 columns)

One row per (window_idx, policy) pair — exactly 8 policy rows per window, using the same 8-policy library as selector_v2_registry: admission_control, best_fit, edf, estimated_service_time_first, fifo, multi_bin_batching, scorpio_style_slo_guard, weighted_shortest_processing. Columns are per-(window, policy) simulator outcome metrics: goodput, latency/TTFT/TPOT/TBT percentiles, queue depths, admission/rejection, and event counts — the identical metric_* schema used by selector_v2_registry, plus one additional domain_id column.

Known research-conclusion limitation — SELECTOR_STATUS = DATA_LIMITED

The experiment that produced this data self-reports SELECTOR_STATUS = DATA_LIMITED. Concretely, on the smallest and final held-out split (FINAL_OOD, n=28 windows, sourced from domains withheld from all model selection), a simple fixed Weighted-Shortest-Processing (WSP) policy baseline achieves lower mean regret (0.0008) than either learned selector variant tested — a domain-balanced random-forest selector (mean regret 0.0016) and a pessimistic-lambda selector (mean regret 0.00096). All three numbers are close and computed over only 28 windows, but as of this data, the learned causal-robustness selector has not been shown to beat a simple fixed baseline in the final held-out domain regime.

What this means for downstream use: this data is suitable for studying window-level policy discriminability, engineering selector features, and comparing candidate selector designs against strong fixed baselines. It does not, by itself, establish that a learned selector generalizes better than WSP to genuinely unseen trace domains. Do not cite this dataset as evidence that a learned selector outperforms simple heuristics on out-of-distribution traces; the authors' own next step is to add more official-source domains before drawing that modeling conclusion.

Structural null columns (not missing data / not corruption)

• selector_v3_policy_vectors: 13 of 52 columns are 100% null across all 20,544 rows (percentile-latency-breakdown, GPU-utilization, and queue-depth-mean/p95 metrics not computed by this simulator run: metric_p50_ttft, metric_p50_tpot, metric_p99_tpot, metric_p50_tbt, metric_p99_tbt, metric_prefill_gpu_utilization, metric_decode_gpu_utilization, metric_prefill_queue_mean, metric_prefill_queue_p95, metric_decode_queue_mean, metric_decode_queue_p95, metric_bridge_queue_mean, metric_bridge_queue_p95). A further 16 columns are null on the same 2,494/20,544 rows (12.14%) — a consistent per-row subset, not scattered missingness.
• selector_v3_windows: 17 of 98 columns are 100% null (disaggregated-serving and multi-instance-migration features — this experiment used monolithic scheduling only, so these features are structurally inapplicable) plus 4 legacy feat_* columns unpopulated by this pipeline version (feat_arrival_rate_recent, feat_arrival_rate_prefix, feat_saturation_load_estimate, feat_recent_slo_violation_rate). Two columns (time_slice_row_start, time_slice_row_end) are null on exactly the 640 synthetic-domain windows, which by construction have no real-trace row range. The complete new 33-column feat_v3_* block is 100% populated (0 nulls across 84,744 cells).

Intended use

Selector-robustness research: training/evaluating policy-selection models under domain shift, studying which engineered features (including the new feat_v3_* rolling-window block) are predictive of policy discriminability, and benchmarking learned selectors against strong fixed baselines.

Provenance

Generated from llm-serving-heuristic-evolution at commit c8aee129f553f8dc3ede99eac60d5b14484beb41.

Source experiment: selector_v3_multidomain_causal_20260721T151341Z. Leakage audit passed (0 duplicate window IDs, 0 group-atomicity violations). See metadata/provenance.json for machine-readable provenance.

Provenance

Generated from llm-serving-heuristic-evolution at commit e8bd759b6cdaa8a05096b0ceeb1c7684cfa07302.

The source experiment was v2_selector_regret_benchmark_20260722T134925Z. Validation manifests passed for dataset construction, leakage audit, model combination, finalist combination, and final report generation. The selected finalist was finalist_direct_multiclass_classification_hist_gradient_boosting_s101.

See metadata/provenance.json and metadata/validation_report.json for machine-readable provenance and validation checks.

Upstream Workload/Trace Attribution

This release contains Soroush-generated derived metrics, features, selector outputs, oracle fields, and regret/suitability records. It does not redistribute raw upstream trace rows, raw prompts, completions, provider payloads, or operational logs.

Rows may be derived from or identify these upstream workload families:

• BurstGPT: public workload trace repository, CC BY 4.0. Source: https://github.com/HPMLL/BurstGPT. If you use rows derived from BurstGPT source families, also cite the BurstGPT dataset/paper as requested by its maintainers.
• Azure LLM Inference Trace 2023: public Azure trace sample, CC BY Attribution License. Source: https://github.com/Azure/AzurePublicDataset/blob/master/AzureLLMInferenceDataset2023.md. If you use rows derived from azure_2023_* or azure_llm_2023 source families, also cite the Azure trace and the Splitwise paper as requested by the dataset provider.
• Synthetic policy frontier / synthetic stress scenarios: generated by Soroush Vahidi's research workflow and covered by this dataset's license for the released metrics.

This dataset does not claim ownership of upstream trace data. No statement in this dataset should be read as claiming ownership of upstream BurstGPT or Azure trace data.

License

License for the released Soroush-generated metrics and derived selector outputs: Creative Commons Attribution 4.0 International (CC BY 4.0).

This license requires attribution. It applies to this dataset's released derived metrics and metadata. It does not replace the licenses or attribution requirements of upstream workload/trace sources.

Attribution

If you use, redistribute, adapt, benchmark with, or build derived artifacts from this dataset, please attribute the dataset to Soroush Vahidi and cite the dataset using the citation below. Upstream workload/trace sources should also be cited as specified in the provenance section.

Citation

@dataset{vahidi_llm_serving_selector_regret_2026,
  author    = {Soroush Vahidi},
  title     = {LLM-Serving Selector Regret},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {1.0},
  url       = {https://huggingface.co/datasets/SoroushVahidi/llm-serving-selector-regret}
}

Limitations

This is a decision-quality/regret benchmark, not a claim that learned top-1 policy selection is solved. Project documentation records that the selected model produces useful suitability/ranking signals but does not fully capture held-out OOD V1-to-V2 oracle gain. Users should analyze split-specific regret and OOD behavior rather than relying only on aggregate averages.

The selector_v3_windows/selector_v3_policy_vectors configs carry their own open research-conclusion caveat: on the smallest held-out split (FINAL_OOD, n=28), a fixed WSP baseline still has lower regret than the learned selectors tested against it (see SELECTOR_STATUS = DATA_LIMITED above). Treat this data as suitable for feature engineering and selector-design comparison, not yet as evidence of a learned selector beating simple heuristics out-of-distribution.