gaia-hazlab/repere-codameter

Repère codameter: dv/v processing tasks (public development split)

Agent-evaluation tasks for ambient-noise relative seismic velocity change (dv/v) monitoring. In param_recommendation the model chooses processing parameters (estimator, frequency band, coda window, stacking, reference); in dvv_series it returns the recovered dv/v(t) series. Both are scored against synthetic data with a known dv/v, generated deterministically by codameter from each case recipe. This release holds the public development split only; the evaluation corpus is withheld.

Part of Repère, an open benchmark for AI agents in the geosciences. Rows follow Repère's BenchmarkRow schema; rebuild a row's scorer with make_scorer_from_spec(row['scorer_spec']) from the suite named in the table below.

Contents

Config	Rows	Task kind	Model must return
param_recommendation	2	code_generation	A JSON processing configuration (estimator, frequency band, coda window, stack length, reference, quality gate).
dvv_series	2	code_generation	The recovered dv/v(t) series, one value per day.

Scoring.

• param_recommendation: Scorer dvv_recovery runs the configuration through codameter on the case's synthetic data and grades the RMS error of the recovered dv/v against the imposed one, from 1 at or below rms_ceiling to 0 at rms_bad. Missing core keys or an invalid band, window or estimator score 0.
• dvv_series: Scorer dvv_series_regression grades the RMS error of the returned series against the imposed dv/v on the same scale (rms_ceiling to rms_bad); rms_bad is set so a no-change series scores about 0.

Only the public development split (validation) is released. The evaluation split is withheld so that ranked scores measure capability rather than recall; scores on this split are for development and are not rankings.

from datasets import load_dataset
ds = load_dataset("gaia-hazlab/repere-codameter", "param_recommendation", split="validation")

Metadata

• croissant.json: Croissant 1.1 with the Responsible AI extension.
• tasks.jsonld: one Croissant Tasks TaskProblem per config (input, output, implementation, execution and evaluation specs). Croissant Tasks is experimental upstream.
• manifest.json: sha256 and row count of every file. A Repère score names the manifest it was computed on.

Dependency

Scoring regenerates each case's data with codameter 0.3.0, commit 9c3259b982ad79b35fb5e85cc8fb0952dc09244a. Install that commit to reproduce scores.

License

mit.

Readiness profile (v0 self-assessment, not reviewed)

Evidence level: 0 missing, 1 present but incomplete, 2 verifiable by another team, n/a not yet assessed. Criteria follow the AI-readiness instrument (A analysis readiness, B AI readiness, D agent-evaluation datasets). Levels are never summed.

ID	Criterion	Level	Evidence
A1	Provenance & rights	2	MIT license; generator pinned at codameter 9c3259b; manifest sha256 per file
A2	Physical comparability	1	dv/v units and sign stated in each prompt; no separate conventions record
A3	Quality & gaps	n/a
A4	Structure & metadata	2	croissant.json (Croissant 1.1 + RAI) and tasks.jsonld (Croissant Tasks)
B1	Diversity	1	independent unit = case recipe; 2 public cases
B7	Independent partitioning	1	cases split by recipe; withheld cases carry secret truth parameters (codameter golden.py)
D1	Task specification	2	scorer_spec on every row; tasks.jsonld
D2	Reference validity	1	dvv_series scorer anchored so a no-change series scores about 0 (codameter); tests/test_dvv_suite.py
D3	Contamination	1	public development split only; withheld split private; no rotation log yet
D4	Task diversity	1	2 task families x 2 public cases (grades easy, medium)
D5	Unanswerable cases	0	no case where abstention is correct
D6	Score stability	1	leaderboard schema records repeats; published board is one run per configuration

Responsible AI

• dataCollection: Synthetic. Each case is a recipe (application, grade, seed, SNR, record length); codameter regenerates daily cross-correlation functions with an imposed dv/v history from the recipe. No field recordings, no human subjects.
• dataCollectionType: Simulation
• dataLimitations: Two development cases (easy volcano, medium earthquake fault zone). The withheld corpus has 30 cases over three grades and six applications (codameter golden.py). Synthetic coda omits source-distribution changes and instrument problems found in field data. Scores on this split are for development, not ranking.
• dataBiases: Cases follow codameter's synthetic model; a method tuned to that model may not transfer to field data.
• personalSensitiveInformation: None.
• dataUseCases: Developing and debugging agents and skills for dv/v processing choices; reproducing the Repère development board. Not for training.
• dataReleaseMaintenancePlan: Versioned with Repère; a changed recipe or scorer is a new version with a new DOI. Withheld cases rotate independently.

Citation

@misc{repere_codameter_2026,
  author = {Denolle, Marine},
  title = {Repère codameter: dv/v processing tasks (public development split)},
  year = {2026},
  publisher = {Hugging Face},
  version = {v0.1},
  url = {https://huggingface.co/datasets/gaia-hazlab/repere-codameter},
}