SoroushVahidi/module-intervention-credit

Module Intervention Credit: Counterfactual Single-Module Substitution Data for LLM-Serving Scheduling Policies

Dataset Purpose

LLM-serving schedulers are typically evaluated and compared as whole, monolithic policies (e.g. FIFO vs. EDF vs. a vendor's proprietary scheduler). This makes it hard to answer a narrower, more actionable question: which internal module of a scheduling policy — its admission control, its priority function, its KV-cache guard, its fairness/aging term, its prefill handling — is actually responsible for a performance difference?

This dataset was built to support that question directly. It contains the results of a controlled counterfactual experiment: for pairs of scheduling policies simulated on the same synthetic workload, one module of a "base" policy is swapped for the corresponding module of a "donor" policy, and the resulting change in reward is recorded as a row. The dataset also includes the downstream attempt to learn a general module-substitution selector from this data via a 10,000-trial AutoML sweep, together with its outcome — which was a negative / weak-generalization result (see the dedicated section below). Both the counterfactual data and the negative modeling result are published, because the negative result is itself informative for anyone attempting similar module-attribution or credit-assignment work in LLM-serving scheduling.

All data was generated by a discrete-event LLM-serving simulator (part of the llm-serving-heuristic-evolution research codebase); it is simulation data, not measurements from a production LLM-serving deployment.

What One Row Represents

intervention_credit (the core table — 8,640 rows)

Each row is one counterfactual module transplant evaluated on one workload configuration:

• base policy (base_policy): the scheduling policy whose module is being replaced.
• donor policy (donor_policy): the scheduling policy that the replacement module is taken from.
• module transplant (module_type, child_name, intervention_id): which single module slot was swapped (admission, priority, kv_guard, fairness_aging, or prefill), a human-readable name for the resulting hybrid ("child") policy, and a stable 12-character hex identifier for the specific intervention definition.
• workload configuration (config_id, regime): which of the 120 synthetic workload configurations (one of 8 regimes) the three policies below were evaluated on.
• intervention reward (reward_intervention): the reward achieved by the hybrid policy (base policy with the donor's module substituted in) on that workload.
• intervention credit: the causal attribution of that swap, expressed as two directional differences — intervention_gain (credit relative to the base policy) and gain_vs_donor (credit relative to the donor policy) — plus several derived pass/fail credit labels (beats_both_parents, expands_envelope, meaningful_expands_envelope, meaningful_gain_001/005/010). See Exact Label Formulas below.

reward_base and reward_donor are the unmodified rewards of the base and donor policies (run as themselves, not as hybrids) on the same workload configuration, included so that intervention_gain and gain_vs_donor are reconstructable without needing to join back to reward_vectors.

reward_vectors (11,880 rows)

The full reward vector this experiment was built from: for every one of the 120 workload configurations, the reward of every evaluated policy on that workload — both the 27 native (whole, unmodified) v2 policies (policy_kind = native_v2) and the 72 module-intervention hybrid ("child") policies (policy_kind = module_intervention_child). 27 + 72 = 99 policies × 120 configurations = 11,880 rows. intervention_credit is a derived/join-friendly slice of this table with credit labels attached; reward_vectors is the underlying raw reward surface.

workload_design (120 rows)

One row per synthetic workload configuration: the regime it belongs to, its random seed, and every generative parameter (arrival process, prompt/output token distribution parameters, SLO tightness, priority skew, resource capacities) used to simulate it.

features (120 rows)

One row per workload configuration (joins 1:1 with workload_design on config_id): a derived, model-ready feature vector summarizing the realized workload trace (queueing statistics, prompt/output token quantiles, SLO slack statistics, priority statistics, per-architecture resource-utilization estimates, and echoed generative parameters). This is the feature table the AutoML sweep in automl_trials was trained against.

intervention_definitions (72 rows)

The specification of every module-substitution intervention: which module slot was swapped, the resulting component-level configuration of the hybrid ("child") policy, and the intervention_id that links each definition to its 120 rows in intervention_credit (72 × 120 = 8,640).

module_registry (27 rows)

A registry of all 27 whole scheduling policies known to the simulator, whether each one's internal modules could be safely decomposed/reconstructed for intervention purposes (reconstruction_status, substitutable), and which of its module slots are marked interchangeable. This registry documents more module slots than this v1 dataset exercises — see Module Coverage below.

reconstruction_fidelity (10 rows)

The fidelity gate results for the 10 policies considered as candidate intervention bases: how closely each policy's decomposed/reconstructed module-based simulation matched its original monolithic implementation, and whether it passed the gate used to decide eligibility as an intervention base. See Fidelity Gate below.

automl_trials (10,000 rows)

One row per AutoML trial from a 10,000-trial overnight hyperparameter/model-family sweep that attempted to learn a general module-credit selector (a model that predicts, from features, which donor-module substitution will help a given workload) from the intervention_credit data. Each row records the trial's hyperparameters, model family, validation metrics, and (for a small labeled preview) held-out test metrics. See Weak-Generalization Result below for why this sweep did not produce a deployment-ready selector, and AutoML Validation vs. Held-Out Test Fields in the data dictionary for how to read this table correctly.

Exact Label Formulas

Verified directly against the published intervention_credit values (float tolerance ~1e-9 unless noted):

intervention_gain = reward_intervention - reward_base
gain_vs_donor      = reward_intervention - reward_donor

beats_both_parents          = 1  iff  intervention_gain > eps  AND  gain_vs_donor > eps
expands_envelope             = 1  iff  (reward_intervention - v2_envelope) > eps
meaningful_expands_envelope  = 1  iff  expands_envelope == 1  AND  (reward_intervention - v2_envelope) > 0.005

meaningful_gain_001 = 1  iff  intervention_gain > 0.001
meaningful_gain_005 = 1  iff  intervention_gain > 0.005
meaningful_gain_010 = 1  iff  intervention_gain > 0.010

eps is a small positive floating-point tolerance (empirically anywhere in [1e-12, 1e-6] is consistent with every published row) used so that reward ties that differ only by floating-point noise are not counted as a "win." Without this tolerance, a small number of exact-tie rows (reward difference on the order of 1e-16) would be misclassified as positive.

v2_envelope is the maximum reward achieved by any of the 27 native (non-hybrid) v2 policies on that workload configuration (i.e. max(reward) over reward_vectors rows with policy_kind == native_v2 for that config_id); v2_best_policy is the (a) policy achieving that maximum — where more than one native policy is tied for the maximum, the recorded name reflects the source pipeline's tie-breaking order rather than a unique winner.

Synthetic Workload Generation

All 120 workload configurations are fully synthetic and parametrically generated — this dataset contains no third-party request traces (no real user prompts, no production traffic logs, no scraped datasets).

• 120 configurations, 120 distinct random seeds (one seed per configuration; see workload_design.seed).
• 8 regimes × 15 configurations per regime: low_load_wsp, moderate_mixed, slo_tight_scorpio, bursty_boundary, prefill_heavy, decode_heavy, kv_pressure, high_noise — each regime fixes a distinct combination of load level, burstiness, SLO tightness, and priority-skew ranges, and each regime is instantiated at 15 different seeded parameter draws.
• Each configuration is generated parametrically: an arrival process (Poisson-like base rate with optional burst multiplier/duration), a lognormal prompt-token-count distribution and a lognormal output-token-count distribution (parameterized by prompt_mean/prompt_cv and output_mean/output_cv respectively), SLO tightness and urgency parameters, a priority skew parameter, and simulated resource capacities (GPU sequence capacity, KV-token budget, step-token budget).
• The realized per-configuration workload trace is then summarized into the features table (queue dynamics, token quantiles, SLO slack statistics, etc.) that downstream modeling (automl_trials) is trained on.

Module Coverage

module_registry documents 27 whole policies and marks, per policy, which of nine module slots (admission, priority, deadline_laxity_guard, prefill, decode, batching_token_budget, placement, kv_guard, fairness_aging) are believed substitutable in principle.

This v1 dataset's intervention_credit rows only exercise five of those nine slots:

admission, priority, kv_guard, fairness_aging, prefill

Do not infer from module_registry alone that all nine registry slots were experimentally intervened upon — the registry describes the simulator's broader substitutability analysis, while intervention_credit covers the subset of slots (five) for which single-module intervention experiments were actually run and published here.

Fidelity Gate

Before a policy could be used as an intervention base in this experiment, its decomposition into independently substitutable modules had to be validated against its original, monolithic (non-decomposed) simulation behavior. This is the reconstruction_fidelity table: for each candidate base policy it reports the mean/max absolute deviation in a normalized workload-goodness metric (mean_abs_anwg_delta, max_abs_anwg_delta) and the fraction of simulation steps where the decomposed policy chose the same first action as the original (mean_first_action_agreement) relative to the monolithic implementation.

Of the 10 policies evaluated for this gate, 7 passed (fidelity_gate = PASS, mapping_status = EXACT_OR_FAITHFUL_PROXY) with near-zero deviation, and 3 failed (fidelity_gate = FAIL, mapping_status = EXCLUDED_FROM_BASES) with non-trivial deviation (least_laxity_first, kv_constrained_online, adaptive_chunked_prefill). Only the 7 passing policies were used as intervention bases in intervention_credit. This gate exists so that a measured "intervention gain" reflects a genuine module-level causal effect rather than an artifact of an imperfect decomposition of the base policy itself.

Weak-Generalization Result (Read Before Modeling)

This section states the verified findings of this project's own downstream modeling attempt, without spin, because it materially affects how this dataset should be used.

• The downstream module-credit selector's generalization status is WEAK_GENERALIZATION. A subsequent, larger overnight retraining effort (the 10,000-trial sweep published here as automl_trials) improved on this but was still assessed as IMPROVED_BUT_NOT_READY, and the associated structural-synthesis capability (automatically composing new policies from learned module credit) was assessed as NOT_READY.
• 10,000 AutoML trials did not produce a held-out top-1 selector that beat both parent policies on the held-out evaluation split. Across the top candidate trials on the labeled test-preview split, top1_beats_both_parents_fraction is 0.0.
• In the underlying intervention_credit data itself, only 42 of 8,640 intervention rows (≈0.49%) satisfy beats_both_parents = 1 — i.e. produce a hybrid policy that strictly beats both its base and its donor. This sparsity is a primary reason the learned selector struggles: there are very few positive examples to learn from, and the positive class is heavily imbalanced.
• The overnight run's own diagnosis (recorded verbatim in provenance/provenance_summary.json's lineage and in the source final_report.md) flags class imbalance, sparse "beats-both-parents" signal, sparse envelope-expansion signal, small effective per-regime sample size, and possible validation/test distribution shift as contributing factors.

This is a limitation of downstream generalization and signal sparsity — evaluated honestly — not evidence that the published rows themselves are corrupted, mislabeled, or unreliable. Every row in intervention_credit and reward_vectors is a directly simulated result, not a model prediction; the negative result concerns only the separate attempt to learn a general selector from that data.

Why publish a negative result: negative/weak-generalization findings are chronically underreported in ML and systems research, which biases the literature and causes redundant effort. Publishing this dataset together with the exact AutoML trial table that produced the negative result lets other researchers (a) reproduce the negative finding exactly, (b) attempt alternative modeling approaches on the same ground truth without re-running the simulator, and (c) calibrate expectations for how hard module-credit generalization is in this setting.

Intended Uses

• Scheduler module attribution: studying which scheduling sub-components (admission, priority, KV guard, fairness/aging, prefill handling) drive performance differences between policies.
• Counterfactual module-substitution analysis in LLM-serving scheduling.
• Intervention-credit modeling: training and evaluating models that predict intervention outcomes from workload features (using features + intervention_credit, or intervention_credit + workload_design directly).
• Policy-composition studies: research into building new scheduling policies by combining modules from existing ones, informed by measured (not assumed) module-level credit.
• Reproducing the documented negative selector result: automl_trials is published specifically so the WEAK_GENERALIZATION / IMPROVED_BUT_NOT_READY finding can be independently checked, extended, or challenged.

Non-Intended Uses

This dataset should not be presented or used as:

• Evidence about real-datacenter LLM-serving workloads — all workloads are synthetic and parametrically generated.
• Proof of causal generalization outside this simulator — findings are specific to the simulated environment and policy set described here.
• Production LLM traffic data of any kind.
• A universal or general-purpose LLM-serving scheduling benchmark.
• Evidence that the learned module-credit selector is deployment-ready — the verified result is the opposite (WEAK_GENERALIZATION / NOT_READY).

How This Dataset Differs from Existing Datasets

A novelty audit was conducted against Hugging Face, Zenodo, GitHub, and the relevant arXiv LLM-serving/scheduling literature prior to preparing this release. That audit did not find a standalone public dataset exposing the same row-level module-substitution intervention-credit representation used here — i.e. rows keyed by (base policy, donor policy, module slot, workload configuration) with directly computed causal-credit labels relative to both parent policies and to a whole-policy performance envelope.

This dataset is distinct from, and does not overlap with, several adjacent categories of existing public resources:

• Raw LLM request/workload traces (e.g. public inference-serving trace datasets): those datasets record observed request timing/token statistics from real or replayed traffic, not counterfactual module-substitution outcomes; this dataset contains no traces at all, only synthetic parametric workload configurations and simulated policy outcomes.
• Scheduler ablation tables reported only in papers: prior scheduling papers often report a handful of ablation numbers in a table or figure, without releasing the underlying per-configuration, per-module row-level data; this dataset publishes the full row-level counterfactual data (8,640 intervention rows × 120 workload configurations) rather than a summary table.
• Reward-model datasets (RLHF-style preference/reward datasets for language generation): those datasets label the quality of language model outputs; this dataset instead labels the outcome of substituting a scheduling policy module, with no relationship to output-quality reward modeling.
• RL / token-level credit-assignment datasets: those typically assign credit to actions or tokens within a single policy's trajectory; this dataset assigns credit to a module substitution between two whole policies, evaluated as a single scalar reward difference per workload configuration, which is a different unit and level of analysis.

This description is deliberately restrained: no claim of being first, largest, unique, comprehensive, or state-of-the-art is made. The claim is narrower — the audit did not locate a public dataset with this specific row-level representation.

Difference from the User's Previous Datasets

This dataset is scientifically distinct from other datasets previously published by the same author:

• SoroushVahidi/lafc-evict — cache-eviction candidate supervision data. Different subsystem (cache eviction, not scheduling/dispatch), different task (candidate supervision, not module-substitution credit).
• SoroushVahidi/lafc-evict-sample — a synthetic publication test sample of the above; not a scheduling dataset.
• SoroushVahidi/scidocs — a third-party information-retrieval benchmark mirror; not original research data and unrelated to LLM-serving scheduling.

module-intervention-credit is a new LLM-serving scheduler-intervention dataset. It is not another version of LAFC-Evict and shares no rows, workloads, or labeling scheme with the datasets above.

Complete Data Dictionary

Dtypes below are the published Arrow/Parquet dtypes. "Role" indicates whether a field is a join key, an input feature, a label, or metadata.

intervention_credit (8,640 rows, 21 columns)

Column	Dtype	Role	Meaning
config_id	string	key	Workload configuration identifier; joins to workload_design, features, reward_vectors.
regime	string	metadata	One of the 8 workload regimes (see Synthetic Workload Generation).
split_group	string	metadata	Internal development/held-out modeling flag (development / heldout) used during original research for model selection. This is not a Hugging Face train/test split — no train/test/validation split is declared for this config; treat split_group as an ordinary categorical column if you use it at all.
intervention_id	string	key	12-character hex identifier of the specific module-substitution intervention; joins to intervention_definitions.
child_name	string	metadata	Human-readable name of the hybrid policy, e.g. admission_control__admission_from__scorpio_style_slo_guard.
base_policy	string	input	Name of the base policy (module donor is being substituted into).
donor_policy	string	input	Name of the donor policy (module source).
module_type	string	input	Which module slot was swapped: one of admission, priority, kv_guard, fairness_aging, prefill.
reward_base	double	input	Reward of the unmodified base policy on this workload configuration.
reward_donor	double	input	Reward of the unmodified donor policy on this workload configuration.
reward_intervention	double	input	Reward of the hybrid (post-substitution) policy on this workload configuration.
intervention_gain	double	label	reward_intervention - reward_base. See Exact Label Formulas.
gain_vs_donor	double	label	reward_intervention - reward_donor. See Exact Label Formulas.
beats_both_parents	int64 (0/1)	label	1 if the hybrid strictly beat both reward_base and reward_donor (tolerance-adjusted). Only 42/8,640 rows (≈0.49%) are 1.
v2_envelope	double	metadata	Max reward among the 27 native (whole) v2 policies on this workload configuration.
v2_best_policy	string	metadata	Name of a native v2 policy achieving v2_envelope on this configuration.
expands_envelope	int64 (0/1)	label	1 if reward_intervention exceeded v2_envelope (tolerance-adjusted) — i.e. the hybrid beat every native whole policy.
meaningful_expands_envelope	int64 (0/1)	label	1 if expands_envelope == 1 and the margin over v2_envelope exceeds 0.005.
meaningful_gain_001	int64 (0/1)	label	1 if intervention_gain > 0.001.
meaningful_gain_005	int64 (0/1)	label	1 if intervention_gain > 0.005.
meaningful_gain_010	int64 (0/1)	label	1 if intervention_gain > 0.010.

reward_vectors (11,880 rows, 4 columns)

Column	Dtype	Role	Meaning
config_id	string	key	Workload configuration identifier. 99 rows per config_id (27 native + 72 hybrid policies).
policy_name	string	key	Name of the evaluated policy (native policy name, or hybrid child_name).
policy_kind	string	metadata	native_v2 (one of the 27 whole policies) or module_intervention_child (one of the 72 hybrids).
reward	double	label	Composite simulated reward/suitability score for this policy on this workload configuration (observed range in this release: ≈0.449–1.0; higher is better).

workload_design (120 rows, 21 columns)

Column	Dtype	Role	Meaning
config_id	string	key	Workload configuration identifier.
regime	string	metadata	One of the 8 workload regimes.
seed	int64	metadata	Random seed used to generate this configuration (range 717171–720860 in this release).
n_requests	int64	input	Number of simulated requests in this workload (54–111).
base_arrival_rate	double	input	Base request arrival rate parameter.
burst_multiplier	double	input	Multiplier applied to arrival rate during a burst.
burst_duration_fraction	double	input	Fraction of the simulation duration spent in a burst state.
interarrival_cv	double	input	Coefficient of variation of inter-arrival times.
prompt_mean	double	input	Mean of the lognormal prompt-token-count distribution.
prompt_cv	double	input	Coefficient of variation of the prompt-token-count distribution.
output_mean	double	input	Mean of the lognormal output-token-count distribution.
output_cv	double	input	Coefficient of variation of the output-token-count distribution.
prediction_noise	double	input	Noise level applied to the policy's output-length prediction signal.
slo_tightness	double	input	SLO deadline tightness parameter (higher = tighter).
urgent_fraction	double	input	Fraction of requests marked urgent.
priority_skew	double	input	Skew parameter for the priority-assignment distribution.
gpu_sequence_capacity	int64	input	Simulated max concurrent sequences per GPU.
kv_token_budget	int64	input	Simulated KV-cache token budget.
step_token_budget	int64	input	Simulated per-step token budget.
split_group	string	metadata	Same internal development/held-out flag as in intervention_credit; not a Hugging Face split.
selection_allowed	bool	metadata	Internal flag from the original research pipeline indicating eligibility for downstream selector evaluation.

features (120 rows, 75 columns)

One row per config_id (joins 1:1 with workload_design). Columns fall into these groups (full column list is in the Parquet schema; grouped here for readability):

Column group	Example columns	Meaning
Arrival / queue dynamics	arrival_rate_recent, arrival_rate_prefix, inter_arrival_cv, burstiness_cv, queue_length, recent_queue_growth_rate, active_sequence_count, saturation_load_estimate	Realized arrival-process and queueing statistics observed in the simulated trace.
Prompt-token statistics	prompt_mean, prompt_median, prompt_p90, prompt_p95, prompt_variance, prompt_cv	Realized prompt-length distribution statistics.
Output-token statistics	pred_output_mean, pred_output_median, pred_output_p90, pred_output_p95, pred_output_variance, pred_output_cv	Realized (predicted) output-length distribution statistics.
SLO / slack statistics	tight_slo_fraction, mean_slack, p10_slack, minimum_slack, recent_slo_violation_rate	Realized SLO-deadline slack and violation statistics.
Priority statistics	priority_mean, priority_p90, priority_high_fraction, priority_class_count	Realized request-priority distribution statistics.
Resource / capacity	resource_gpu_count, resource_kv_capacity, resource_block_size, resource_sequence_capacity, resource_token_budget	Simulated resource capacities.
Monolithic-architecture utilization	monolithic_aggregate_kv_utilization, monolithic_active_batch_size	Utilization estimates under a monolithic (non-disaggregated) serving architecture.
Disaggregated-architecture fields	disagg_prefill_gpu_count, disagg_decode_gpu_count, disagg_prefill_queue_length, disagg_decode_queue_length, disagg_bridge_queue_length, disagg_prefill_side_utilization, disagg_decode_side_utilization, disagg_transfer_delay_s	Utilization/queueing estimates under a disaggregated prefill/decode serving architecture.
Multi-instance fields	multi_instance_count, multi_instance_load_imbalance, multi_instance_kv_imbalance, multi_instance_incoming_migration_count, multi_instance_migration_pressure	Estimates for a multi-instance serving deployment (load balancing / migration pressure).
Derived summary fields	derived_duration, derived_offered_request_rate, derived_prompt_p95, derived_output_p95, derived_prefill_decode_ratio, derived_approx_load_ratio	Derived scalar summaries of the workload trace.
Echoed generative parameters	param_seed, param_n_requests, param_base_arrival_rate, param_burst_multiplier, param_burst_duration_fraction, param_interarrival_cv, param_prompt_mean, param_prompt_cv, param_output_mean, param_output_cv, param_prediction_noise, param_slo_tightness, param_urgent_fraction, param_priority_skew, param_gpu_sequence_capacity, param_kv_token_budget, param_step_token_budget, param_selection_allowed	The generative parameters from workload_design, echoed here (same values) for modeling convenience so features is self-contained without a join.
config_id, regime	—	key / metadata

All numeric feature columns are double; boolean-valued echoed parameters (e.g. param_selection_allowed) are bool; config_id/regime are string.

intervention_definitions (72 rows, 12 columns)

Column	Dtype	Role	Meaning
name	string	metadata	Same value as child_name in intervention_credit.
admission	string	input	Name of the admission-control component used in this hybrid (present when module_type = admission, else the base policy's own component).
priority	string	input	Name of the priority component used in this hybrid.
placement	string	input	Name of the placement component used in this hybrid (not an intervenable slot in v1; carried through from the base/donor definition).
base_policy	string	input	Base policy for this intervention.
donor_policy	string	input	Donor policy for this intervention.
module_type	string	input	Which slot this intervention swaps: admission, priority, kv_guard, fairness_aging, or prefill.
donor_module	string	input	Name of the specific donor component substituted in.
intervention_id	string	key	12-character hex identifier; joins to intervention_credit.intervention_id.
kv_guard	string	input	Name of the KV-guard component used in this hybrid (null unless module_type = kv_guard).
fairness_aging	string	input	Name of the fairness/aging component used in this hybrid (null unless module_type = fairness_aging).
prefill	string	input	Name of the prefill component used in this hybrid (null unless module_type = prefill).

module_registry (27 rows, 23 columns)

Column	Dtype	Role	Meaning
policy_name	string	key	One of the 27 whole scheduling policies known to the simulator.
library	string	metadata	Which policy library/generation the policy belongs to (v1 or v2_new).
reconstruction_status	string	metadata	Whether this policy's modules could be decomposed for intervention: EXACT_CANDIDATE, PARTIAL, or UNSUPPORTED.
substitutable	bool	metadata	Whether any module substitution was considered valid for this policy.
unsupported_reason	string	metadata	Free-text reason when substitutable = False or reconstruction is only PARTIAL.
admission, priority, deadline_laxity_guard, prefill, decode, batching_token_budget, placement, kv_guard, fairness_aging	string	metadata	Name of this policy's implementation of each of the nine module slots.
admission_substitutable, priority_substitutable, deadline_laxity_guard_substitutable, prefill_substitutable, decode_substitutable, batching_token_budget_substitutable, placement_substitutable, kv_guard_substitutable, fairness_aging_substitutable	bool	metadata	Per-slot flag: whether the simulator's analysis considers that module slot substitutable for this policy. Note: only admission, priority, kv_guard, fairness_aging, and prefill are actually exercised in this v1 dataset's intervention_credit rows (see Module Coverage).

reconstruction_fidelity (10 rows, 6 columns)

Column	Dtype	Role	Meaning
policy_name	string	key	Candidate base policy evaluated for the fidelity gate.
mean_abs_anwg_delta	double	metadata	Mean absolute deviation, in a normalized workload-goodness metric, between the decomposed and monolithic simulation of this policy.
max_abs_anwg_delta	double	metadata	Max absolute deviation of the same metric.
mean_first_action_agreement	double	metadata	Fraction of decision points where the decomposed policy's first chosen action matched the monolithic implementation's.
fidelity_gate	string	label	PASS or FAIL. Only PASS policies were used as intervention bases.
mapping_status	string	metadata	EXACT_OR_FAITHFUL_PROXY (passed) or EXCLUDED_FROM_BASES (failed).

automl_trials (10,000 rows, 120 columns)

The sanitized 10,000-trial AutoML leaderboard. The raw source table's model_path column (an absolute local filesystem path to a serialized model binary) and a params.trial_id column (verified to be an exact duplicate of trial_id) were removed; no serialized model binaries are published. All 118 remaining metric/hyperparameter columns are unchanged.

Column (or prefix)	Dtype	Role	Meaning
trial_id	int64	key	Canonical trial identifier (0–9,999 range, not necessarily contiguous after AutoML pruning).
status	string	metadata	Trial completion status (all 10,000 published rows are OK).
objective_value	double	label	The AutoML sweep's scalar objective for this trial (used to rank trials).
elapsed_s	double	metadata	Trial wall-clock runtime in seconds.
params.kind	string	input	Model family, e.g. pairwise_logistic, extra_trees_clf, extra_trees_reg, rf_clf, rf_reg, hgb_clf, hgb_reg, xgb_clf, xgb_reg, lgbm_clf, lgbm_reg, torch_mlp_multitask, torch_siamese_multitask.
params.task	string	input	Modeling task type, e.g. ranking.
params.target	string	input	Regression/classification target column name from intervention_credit this trial was trained to predict (e.g. C_base-style credit target); null for trials that don't use a single scalar target (e.g. multitask/ranking trials).
params.seed	int64	input	Random seed for this trial.
params.feature_mode	string	input	Which feature subset from features was used.
params.class_weight, params.n_estimators, params.max_depth, params.min_samples_leaf, params.learning_rate, params.l2, params.dropout, params.hidden, params.batch_size, params.calibration, params.sampling, params.label	mixed	input	Model-family-specific hyperparameters (populated only for the model families that use them; null otherwise).
validation.*	double / int64	label	Validation-split metrics — this is the metric family the AutoML sweep actually used to select trials. Includes top-k selection metrics (top1/top3/top5) such as mean_C_base, mean_C_parent, mean_C_env, positive_precision, meaningful_001_precision, meaningful_005_precision, beats_both_parents_fraction, expands_envelope_fraction, negative_transfer_rate, mean_uncertainty, plus calibration/regression diagnostics (mae, rmse, bias, sign_accuracy, ece, brier, roc_auc) and precision/coverage at several confidence percentiles (precision_at_p50...p90, coverage_at_p50...p90).
test_preview_not_for_selection.*	double / int64	metadata (diagnostic only)	The same metric family as validation.*, computed on a small labeled held-out preview split. The not_for_selection suffix is load-bearing: these columns were explicitly excluded from AutoML trial/model selection and exist only as a diagnostic preview of held-out behavior — do not use them to pick a "best trial" after the fact, as that would reintroduce test-set leakage the original pipeline was designed to avoid.
decision_rule.*	mixed	metadata	The specific acceptance rule (objective, threshold, uncertainty threshold) applied to this trial's predictions, and the resulting decision-level metrics (n_accepted, coverage, mean_C_base, etc.) under that rule.

License

• Code (the llm-serving-heuristic-evolution repository that generated this data): MIT License.
• Dataset (this Hugging Face dataset, module-intervention-credit): CC BY 4.0.

These are two separate licenses for two separate things: the license on the code repository does not apply to this dataset's content, and vice versa. No repository or legal evidence was found during preparation of this release that would make CC BY 4.0 unsafe or inconsistent for the dataset content; if you are the rights holder and believe otherwise, do not publish under this license without review.

Citation

Version and immutable reference

This is dataset version v1, distributed through the canonical Hugging Face repository: https://huggingface.co/datasets/SoroushVahidi/module-intervention-credit. No DOI currently exists for this dataset, and no DOI is claimed. For an immutable reference, cite Hugging Face release revision 5d2469b573aa2d75a69d4d077af6938f002a6c19 and preserve that exact revision. The scientific source commits listed below identify the generation inputs, not the Hugging Face dataset revision.

No verified Soroush Vahidi-authored paper or preprint directly associated with module intervention credit, LLM-serving scheduling, heuristic evolution, or compositional scheduler design is currently known. The source repository is therefore the primary research provenance; no paper relationship is invented.

No verified paper/preprint DOI exists specifically for this dataset, and none is invented here. No Zenodo DOI is invented either — this dataset is currently intended for Hugging Face distribution only.

@misc{vahidi_module_intervention_credit_2026,
  author       = {Vahidi, Soroush},
  title        = {Module Intervention Credit: Counterfactual Single-Module Substitution Data for LLM-Serving Scheduling Policies},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {v1},
  howpublished = {\url{https://huggingface.co/datasets/SoroushVahidi/module-intervention-credit}},
  note         = {Generated from the llm-serving-heuristic-evolution repository, https://github.com/SoroushVahidi/llm-serving-heuristic-evolution; cite the immutable Hugging Face revision recorded in CITATION.cff}
}

Affiliation: Soroush Vahidi, New Jersey Institute of Technology.

Reproducibility

• Intervention / scientific-table source commit: e8bd759b6cdaa8a05096b0ceeb1c7684cfa07302 (branch wulver-final-integration-20260721) in llm-serving-heuristic-evolution. This commit generated intervention_credit, reward_vectors, workload_design, features, intervention_definitions, module_registry, and reconstruction_fidelity.
• AutoML run source commit: aa8b418cf3f7d810776ea815593019a7352075da (branch wulver-selector-v2-and-composition-integrated), which generated automl_trials (seed 20260721, 10,000-trial budget, ~26,571s runtime).
• The repository's current HEAD has advanced past both of these commits. Reproducing this exact dataset requires checking out the recorded historical commit above, not building from current HEAD.

Safe regeneration outline (no private HPC paths required — substitute your own local clone path and output directory):

git clone https://github.com/SoroushVahidi/llm-serving-heuristic-evolution.git
cd llm-serving-heuristic-evolution
git checkout e8bd759b6cdaa8a05096b0ceeb1c7684cfa07302
# Follow the repository's own module-intervention-credit generation pipeline
# (see docs/ and scripts/ at this commit) to regenerate:
#   combined/single_module_credit.csv, combined/reward_vectors.csv,
#   design/workload_design.csv, combined/features.csv,
#   design/intervention_definitions.csv, diagnostics/module_registry.csv,
#   fidelity/reconstruction_fidelity.csv

git checkout aa8b418cf3f7d810776ea815593019a7352075da
# Follow the repository's AutoML module-credit sweep entry point at this commit,
# with seed=20260721 and max_trials=10000, to regenerate leaderboard.csv.

This dataset's own transformation from those source CSVs to the published Parquet files (column selection for automl_trials, dtype normalization, provenance redaction) is versioned as transformation_version: module-intervention-credit-v1-release-prep-1 in provenance/provenance_summary.json.