aashay96/scientific-posttrain-battery-pbt-eval

ScientificPostTrain Battery PBT Evaluation v1

This repository distributes the compact, immutable evaluation artifact for the battery-pbt Harbor benchmark. Starting from the fixed official Pretrained Battery Transformer (PBT) checkpoint, an agent must improve early prediction of equivalent-full-cycle lifetime for real industrial lithium-ion pouch cells. The benchmark is intentionally queryable: agents may repeatedly run the exact evaluation while curating eligible external training data, but these evaluation records must not be used as training supervision.

Files

• battery_pbt_eval_v1.npz: allow_pickle=False NumPy archive used by the agent-visible evaluator and Harbor verifier.
• manifest.json: source hashes, pinned preprocessing revisions, retained-cell audit, cycle policy, and exclusions.
• SHA256SUMS: release-file integrity hashes.

The NPZ contains:

Array	Meaning
curves	Float32 [cells, 50, 3, 300] early-cycle curves. Channels are normalized voltage, C-rate, and normalized capacity.
curve_mask	UInt8 [cells, 50] left-packed valid-cycle mask.
life	Float32 equivalent-full-cycle lifetime target.
group_id	Material-design/protocol group used for macro averaging.
cell_id	Stable source-cell identifier.
cell_type	Published pouch-cell type.

Source And Processing

The source is Zenodo record 17654407, "Discovery Learning predicts battery cycle life from minimal experiments," released under CC BY 4.0. The release contains 123 large-format pouch cells in 37 material-design/protocol groups. The audited evaluation retains 116 cells across 33 groups and accounts for all seven exclusions in manifest.json.

The reproducible builder verifies the Zenodo archive and workbook checksums, uses pinned BatteryLife and BatteryML revisions, excludes RPT segments, retains the first 20-50 valid chronological cycling segments, and transforms each cycle according to PBT's released representation. The manifest records every retained cell and exclusion. The authoring source lives in ScientificPostTrainBench.

Evaluation Contract

For prediction p_i, target y_i, and source group g:

e_i = |log(clip(p_i, 1, 100000)) - log(y_i)|
E_g = mean(e_i for cells i in group g)
E = mean(E_g over represented groups)
scientific_score = exp(-E)

Harbor downloads these files from an immutable Hub commit and verifies their SHA-256 digests. It then strict-loads only the submitted fixed-schema PBT Safetensors and evaluates all retained cells through the task's public /app/evaluate.py.

Intended Use

This artifact is for evaluation and aggregate model selection in the ScientificPostTrainBench task. It is not an agent training set. Evaluation curves, labels, group IDs, cell types, or derivatives must not be fitted, pseudo-labeled, copied into training data, or used for example-level selection. The complete Harbor trajectory and submitted method summary are used for provenance auditing.

The benchmark is not yet claimed to be frontier-calibrated. That designation requires battery-expert review plus replicated base, legal-adaptation, and privileged-target-supervised anchors with uncertainty.

Citation

Please cite the source dataset and PBT publications listed in the benchmark specification, and attribute this derived evaluation package to ScientificPostTrainBench.