Lance1573/route-anything-data

Route Anything: experimental data

Recorded inputs and per-task outputs for the experiments accompanying Route Anything: A Theory of Inference Routers. The code repository is InfRec-Cardinal/route-anything.

These files were collected or computed in September 2026. Packaging them did not run new inference. Outcome-matrix replay, recorded online model calls, and constructed memory streams are separate forms of evidence. The dataset preserves the published task splits, model names, scores, and token-price estimates.

Contents

Path	Population	Purpose
data/agent-as-a-router/data/coderouterbench/	9,999 ID tasks × 8 models; 176 OOD tasks × 8 models	Benchmark outcomes, costs, splits, and task metadata
data/agent-as-a-router/data/id/voter_decisions/	Nine released policies on 2,919 ID Test tasks	Ex-ante model choices for routing-signal analyses
data/agent-as-a-router/outputs/	Released ID choices, OOD traces, and baseline choices	Policy and escalation replay
data/agent-as-a-router/data/matrices/phase1_acrouter_v2/response_matrix.json	Recorded ID model responses	Optional local verifier reconstruction
experiments_x/X7b_inloop_live/	2,317 attempt records on 283 tasks; 849 reference-verifier records	Retry and feedback replay, including separate truncation views
experiments_x/X8_operating_point/	500 probing and 334 runnable held-out tasks	Verifier scores, call usage, samples, and cached judgments
experiments_x/X12b_llm_signal/	600 probing and 600 held-out tasks	Paid forecasts and the associated backend outcomes
experiments_x/X12c_jev_policy/	3,095 official choices: 2,919 ID Test + 176 OOD	Jev direct-choice replay; separate two-primary records retained
experiments_x/X13_memory/cache/	23,288 task-model verdicts on 2,911 tasks	Real verifier outcomes used by constructed memory processes
extensions/paid_forecast/	Original 600/600 forecast sample and two later 1,200-call runs	September 24–25 follow-up: native LoRA actions and base-model probability forecasts
experiments_x/R6_stepwise/results/	Two Qwen3 pairs on 541 EvalPlus tasks	Per-continuation scores, latency fits, and recorded analysis summaries

inventory.json lists each data file, its source, size, and row count where applicable. There are no model weights in this release.

Data layout and use

Download this dataset into a separate directory, then copy its data/, experiments_x/, and optional extensions/ contents into the matching directories of a code checkout. The paths are designed to merge with the repository without replacing its README or Python scripts. Alternatively, set the data locations explicitly:

export ACROUTER_DATA=/path/to/dataset/data/agent-as-a-router/data
export ACROUTER_REPO=/path/to/dataset/data/agent-as-a-router
export ACROUTER_OUTPUTS=/path/to/dataset/data/agent-as-a-router/outputs
export ACROUTER_RELEASE_DECISIONS="$ACROUTER_OUTPUTS/current/id_decisions.jsonl"
export ACROUTER_ID_DECISIONS="$ACROUTER_RELEASE_DECISIONS"

Scripts that use per-experiment caches expect those records under the matching experiments_x/ directory in the code checkout. Use the repository's documented replay entry points to reuse the records. Fresh collection is a separate operation with its own model and service requirements.

The Hugging Face viewer exposes four tabular configurations. paid_forecast retains its original split column with values probing and id_test; it is not a random re-split. memory_verdicts retains task identifiers and obtains its probing/test membership by joining to CodeRouterBench.

Schemas

• Benchmark results: task_id, split, dimension, model, score, cost_usd, token usage, latency, and cost_source. OOD tables use resolved, apply_ok, and the original old112/new64 source split.
• Policy choices: task_id and chosen_model; escalation records additionally retain the reached chain_run and observed costs/outcomes.
• Retry attempts: job_id, task_id, root_model, branch, k, attempted model, token usage, costs, truth and visible-test scores, response text, and extracted code. Separate view files are alternative analysis views, not additional independent tasks.
• Verifier rows: one task per row; columns name each primary's released outcome, cost, latency, and V1/V2/V3 verifier score. Cache records preserve per-task judgments and sample membership.
• Forecasts: task_id, original split, p_qwen, p_kimi, measured usage, several stated price assumptions, backend score/cost columns, and TF-IDF comparison probabilities.
• Memory verdicts: task_id, model, source, released truth, visible-test acceptance (w1_*), and hidden-test outcomes (w2_*). Arrival and retention processes are generated by the experiment code rather than recorded deployment logs.
• Jev choices: task_id, selected choice, returned answer, usage, latency, error, and ID/OOD membership.

Operational cache identifiers, wall-clock timestamps on attempt records, and account-level ledger snapshots were omitted. Scores, costs, token counts, task identifiers, and protocol labels used by the analyses are preserved.

Provenance and terms

The compact benchmark tables and Agent-as-a-Router decision/response artifacts come from LanceZPF/agent-as-a-router. CodeRouterBench is also released at Lance1573/CodeRouterBench. Both the repository license and its dataset card identify MIT terms; the original license text is included in licenses/agent-as-a-router-MIT.txt. Cite Agent-as-a-Router: Agentic Model Routing for Coding Tasks, arXiv:2606.22902, when using these inputs.

Route Anything's experiment records and derived measurements are released with the associated code. Underlying benchmark task text and model-generated code retain their upstream terms. The relevant source benchmarks include HumanEval, MBPP, HumanEval-X, BigCodeBench, LiveCodeBench, DS-1000, CRUXEval, and SWE-derived tasks; consult the upstream benchmark release for task-specific provenance. The R6 scores use HumanEval+ and MBPP+ from EvalPlus, excluding HumanEval/32 under the recorded sandbox protocol.

RouterBench is a separate upstream dataset. Its matrix is not redistributed here because the available local dataset card does not state redistribution terms. Obtain the 0-shot release linked by the RouterBench project, place it at data/routerbench/routerbench_0shot.pkl, and set ROUTERBENCH_PKL if it is stored elsewhere. The paper uses 36,497 prompts and 11 models.

Interpretation and reproduction limits

• The CodeRouterBench split is 7,080 probing and 2,919 ID Test tasks. Retain those splits for fitting and evaluation. Missing recorded runs include 148 zero-score/zero-cost Qwen3-Max rows; zero recorded cost is not a claim that a real call would be free.
• Prices are the experiment's token-price estimates. Local verifier seconds and hosted-model dollar estimates are different resource bases. Some verifier analyses use a stated proxy price.
• X12-c evaluates direct backend choice. It does not evaluate Jev as a post-output verifier. The later Qwen LoRA action run and base-model probability run are different protocols.
• X13 uses real responses and verifier outputs with constructed arrival/write/retention processes. It is a semi-synthetic replay, not a measured production memory trace.
• Reconstructing prompts or executing benchmark tests may require the original benchmark snapshots and dependencies. Full shared service caches and provider account records are not included.
• R6 includes scores and final summaries but lacks the original generated token trajectories. Recomputing its complete token-cost and handoff analysis requires regenerating those trajectories. The archived scores alone do not provide a complete rerun of that study.

See the code repository's experiment guide for commands and the exact scope of each verified replay.