AgentsSci/IEEE2026_BigData_MAS-4-Science-Matching

SciAgentTrace

An execution-layer trace resource for scientific-agent workload characterization.

A protocol fixes who reasons, what each role can see, when feedback returns, and when a workflow stops. Those choices determine the sequence of model requests that produces an answer, so protocol design is also workload design. Two workflows that consume similar token totals can issue very different request sequences. SciAgentTrace records that difference.

The matched core runs the same 6,653 problems from five benchmark families under four collaboration protocols and two actor models — 53,224 trajectories in 40 actor-family-protocol cells. Each trajectory links an ordered trace to an evaluated answer and to an execution record.

from datasets import load_dataset

repo = "AgentsSci/IEEE2026_BigData_MAS-4-Science-Matching"
traj    = load_dataset(repo, "trajectories",     split="test")   # 68,892 rows
matched = load_dataset(repo, "matched_outcomes", split="test")   # 17,223 rows

What makes this different

Most agent trace collections record answers and token totals. This one adds an audited execution layer:

Measurement	Trajectories	Scope
Client-observed elapsed time	53,224	all four protocols
Tokens + observable request counts	39,918	Baseline, Single-Agent, Broadcast
Elapsed time only	13,306	Planner-Executor-Reviewer

Unavailable fields are published as unavailable, never imputed. The 13,306 elapsed-only rows are an audited coverage boundary with a manifest trail, not missing data: that protocol's historical collector mixed worker-cumulative main-model counters with memory-summary counters that reset per problem, and the components cannot be separated exactly after the fact. We withhold the fields rather than estimate them. Those rows carry a large share of the corpus's measured execution time, so this boundary is material and is stated rather than hidden.

cost_coverage_ratio measures completeness only for the fields named in each run's available-field list; a value of 1.0 on a wall-time-only run does not imply complete token accounting.

Composition

Family	Domain	Problems	Trajectories
OmniMath2	competition mathematics	4,181	33,448
JEEBench	pre-engineering science	515	4,120
SciBench	college science	574	4,592
LAB-Bench	biology	741	5,928
MaScQA	materials science	642	5,136
Matched core		6,653	53,224

Actor models: Gemma 4 31B and GPT-OSS 120B. GPT-OSS 120B is the fixed answer evaluator for both arms, at temperature 0, comparing a submitted answer against evaluator-only reference material.

The package also contains 15,668 additional slice-specific trajectories (a 1,542-problem LAB-Bench set and an 833-problem Gemma 3 27B OmniMath2 sample), bringing the trajectory table to 68,892 rows. These carry distinct slice and comparability identifiers and are not pooled into the matched denominator. They have no execution-layer coverage.

Protocols

ID	Name	Structure
baseline_llm	Direct Baseline	One attempt; submit the extracted answer once.
single_agent	Single-Agent Iterative	Up to three attempts; prior submission and compact evaluator feedback retained.
per	Planner-Executor-Reviewer	Up to three outer rounds; Reviewer emits a diagnosis and exactly one route.
broadcast	Broadcast Deliberation	Three peers over a shared candidate; submit on unanimity or budget.

No acting role ever sees the reference answer; only the external evaluator does. An outer round may contain several role messages and, for PER, more than one evaluator call, so round count is not a request-count proxy.

Tables

Config	Rows	Unit
trajectories	68,892	one protocol execution on one problem
outcomes	68,892	one observed final-answer label
matched_outcomes	17,223	one problem/model group across four protocols
problems	7,467	one source problem after duplicate collapse
messages	1,889,265	one ordered trace event

Stable identifiers join the layers without relying on row order: events and outcomes join by trajectory_uid, execution information by run_id, and cross-protocol comparison requires equality of problem_uid.

How to use it correctly

1. Select one benchmark slice and comparability group.
2. Require equality of problem_uid across the protocols you compare.
3. Name the actor and evaluator explicitly.
4. Match the execution configuration (serving system + software version) before comparing elapsed time. Elapsed time is not interpretable across configurations — one Broadcast cell contains rows from a second backend whose median elapsed time differs by roughly 3x from the rest of that cell.
5. Preserve unavailable fields; do not coerce them to zero.
6. Verify the release manifest and checksums rather than matching by row order.

Two anomalies are retained and labeled rather than silently dropped: 24 synchronized-delay executions in one cell, and the backend split noted above.

Measurement boundaries

wall_time_seconds spans the complete per-problem application call, including inference, client orchestration, visible waiting and retries, and final evaluation. It is client-observed elapsed time — not scheduler walltime, request service time, queue time, accelerator time, energy, or dollar cost. Request counts include evaluation and memory summarization; evaluator invocations are a subset of those submissions, not an additional count. Failed requests that returned no usage object, and unobserved lower-level retries, are not counted.

Problem-bootstrap intervals computed on this corpus describe variation across problems within the recorded executions — not repeated runs or service-load states. Controlled performance or capacity claims require randomized repeated runs with serving-system instrumentation, which this resource does not provide.

Licensing

This is a composite resource and the root license is other. No single license governs the whole package. Every row carries its source license and registry/licenses.parquet is authoritative.

• LAB-Bench material is CC-BY-SA-4.0 and carries an upstream contamination canary and a do-not-train request; both pass downstream.
• MaScQA material is CC-BY-NC-SA-4.0 — not licensed for commercial use.
• Other benchmark material remains governed by its upstream terms.

Users must preserve benchmark licenses and attribution, and must not treat generated text as expert advice or as authoritative scientific explanation. Benchmark contamination is unknown.

Citation

A paper describing this resource is under submission to IEEE BigData 2026. Citation details will be added on acceptance. Please also cite the five upstream benchmarks you use.

@misc{sciagenttrace2026,
  title  = {SciAgentTrace: An Execution-Layer Trace Resource for
            Scientific-Agent Workload Characterization},
  author = {Yang, Chih-Hsuan and Thakur, Rajeev},
  year   = {2026},
  note   = {Under submission},
  url    = {https://huggingface.co/datasets/AgentsSci/IEEE2026_BigData_MAS-4-Science-Matching}
}