Allanatrix/ssb_pilots_combined

SSB Pilots Combined (Sanitized)

This dataset packages two production-style pilot outputs for solid-state battery (SSB) candidate screening:

• inference_pilot/: one full scoring run across a mixed candidate pool.
• tuning_pilot/: multi-run tuning and repeatability artifacts used to evaluate ranking stability and lift.

The goal is prioritization, not direct property prediction. Use this to shortlist candidates for downstream DFT/experimental validation when budget is limited.

Why this matters

Materials discovery pipelines usually face extreme class imbalance and tight validation capacity. This dataset shows how ranking-based screening can increase hit density in top-K selections versus random sampling, with explicit uncertainty and repeatability context.

What is included

1) Inference pilot (inference_pilot/)

Core files:

• mixed_inference_input.jsonl: input candidates with candidate_id, composition, and features.
• ranked_candidates.csv: scored and ranked output table (primary file for downstream usage).
• evaluation_lookup.csv: labeled lookup used for post-hoc evaluation on the reference subset.
• top_005permille.csv, top_010permille.csv, top_050permille.csv, top_100permille.csv: precomputed shortlists.
• screening_metrics.json: machine-readable metrics and confidence sweeps.
• screening_report.md: human-readable report.

Inference pilot context (from screening_metrics.json):

• Run timestamp: 2026-02-01T23:00:49.099228+00:00
• Total scored: 33,000
• Eval subset size: 13,000
• Positives in eval subset: 19 (prevalence 0.00146)
• Throughput: 35,645 candidates/s (CPU inference in this pilot)

Top-K enrichment highlights on labeled eval subset:

• Top 0.5% (k=165): enrichment 40.25x
• Top 1% (k=330): precision@k 0.0241, recall@k 0.1053, enrichment 16.49x
• Top 5% (k=1650): enrichment 5.60x
• Top 10% (k=3300): enrichment 3.43x

2) Tuning pilot (tuning_pilot/)

Contains multiple timestamped run folders (for example ssb_tuning_20260201T234218Z) with:

• screening_card.md and screening_card.json: run-level summary and recommended policy.
• repeatability_summary.json: multi-trial aggregate metrics.
• trial_*/ and optional flagship/: per-trial scored outputs, shortlists, and metric summaries.

Flagship tuning snapshot (ssb_tuning_20260201T234218Z/screening_card.json):

• Screened N: 113,000
• Top 1% lift: 11.54x ± 0.99
• Top 5% lift: 7.56x ± 0.24
• Expected hits per 50 experiments: 28.4
• Repeatability (Jaccard mean): top1 0.207, top5 0.210

Data schema (primary table)

inference_pilot/ranked_candidates.csv columns:

• candidate_id: unique candidate identifier.
• composition: composition string.
• predicted_class: numeric class ID from classifier.
• predicted_label: human-readable class label.
• confidence_score: max predicted class probability.
• positive_probability: probability for positive screening target.
• rank: 1-based rank (lower is better).
• rank_percentile: normalized rank in [0, 1].
• is_eval_ref: whether candidate belongs to labeled evaluation subset.
• true_label: reference label (mostly for eval subset).
• tier: source tier metadata.

Recommended usage workflow

1. Load inference_pilot/ranked_candidates.csv.
2. Select a top-K budget (for example top 1% or top 5%).
3. Optionally apply a confidence threshold guided by inference_pilot/screening_metrics.json.
4. Pull diversity-aware alternatives from tuning pilot shortlists if execution constraints require broader chemistry coverage.
5. Send shortlisted candidates to downstream simulation or experimental validation.

Minimal example (Python)

import pandas as pd

ranked = pd.read_csv("inference_pilot/ranked_candidates.csv")

# Example: top 1%
k = max(1, int(0.01 * len(ranked)))
shortlist = ranked.nsmallest(k, "rank")

print(shortlist[["candidate_id", "composition", "positive_probability", "confidence_score"]].head(10))

How to interpret metrics correctly

• lift/enrichment is measured against random selection at the same cutoff.
• Reported precision/recall values are computed on the labeled reference subset, not guaranteed for every external pool.
• Class prevalence can shift strongly across new candidate pools; re-benchmark on your own labeled slice before operational deployment.

Sanitization and privacy

This public release is sanitized:

• Removed all *.log files.
• Removed run_manifest.json files.
• Retained only structured artifacts required for analysis (.csv, .json, .jsonl, .md, .parquet).

Intended use

• Candidate ranking and triage.
• Policy comparison for top-K vs confidence-threshold filtering.
• Repeatability analysis of screening behavior across runs.

Out-of-scope use

• Direct claims of experimental property prediction without further validation.
• Safety-critical decisions without independent verification.

Provenance

• Upstream dataset reference: Allanatrix/SSB_Dataset
• Upstream model reference: Allanatrix/SSB_Screening_Model

Limitations

• Labels are proxy labels for screening, not full experimental truth.
• Performance may degrade under domain shift.
• Repeatability metrics summarize pilot conditions and may differ under different feature pipelines or candidate distributions.