---
name: ocx25-data
description: The OCx25 electrocatalysis dataset (Meta FAIR, University of Toronto, VSParticle) as bound in this Campaign - what each file and column means, the fixed modelling protocol (canonical targets, leave-one-composition-out folds, Pearson R2 of pooled out-of-fold predictions), the published baselines to beat, the rules that stop evaluation drift, and two standard-library scripts that derive the targets and score predictions.
---

# OCx25 data and protocol

Every Problem in this Campaign binds the same three CSV files. Read them
from the Resource URIs `start_attempt` returns; do not fetch another copy.

| Key | What it is |
|---|---|
| `experiments` | Raw testing data, one row per (sample, replicate, reaction, current density). Columns: `sample id`, `source` (`vsp` = VSParticle spark ablation, `uoft` = University of Toronto chemical reduction), `batch number`, `batch date`, `composition` (the *target* composition, `Ag-0.5-Au-0.5` notation), `rep` (`1`,`2`,… or `std`, or `avg`/`std` summary rows), `reaction` (`HER` or `CO2R`), `current density` (mA/cm²; HER always 50, CO2R 50/100/150/200/300), `voltage` (full-cell, V), `fe_h2`, `fe_co`, `fe_ch4`, `fe_c2h4`, `fe_gas_total`, `fe_liquid` (Faradaic efficiency, %; liquid is the electron-balance remainder, assumed formate). |
| `xrf` | Measured compositions per sample: `xrf-gde_formula` (on the electrode; used for `vsp`), `xrf-wafer_formula` (on the wafer; used for `uoft`, keyed by the sample id without `_repN`), per-element means and standard deviations, and occasional LIBS or post-anneal measurements. |
| `dashboard` | The dashboard's cleaned join: experiments with `target composition`, `xrf composition` and one atomic-fraction column per element. Convenient for exploration; the canonical source is `experiments` + `xrf`. |

`reference.md` beside this file has the full data dictionary, the
experimental conditions, the published baseline numbers and the noise
floors. Read it before modelling.

## The fixed protocol

The point of this Campaign is a number that means the same thing in every
Submission. Three things are therefore fixed and not yours to change:

1. **Targets** come from `prepare_targets.py` run on the bound bytes:

   ```
   python3 prepare_targets.py --experiments experiments.csv --xrf xrf.csv --out-dir targets/
   ```

   `targets_her.csv`: one row per synthesis target (sample id without
   `_repN`, per source): `voltage_cell_v` at 50 mA/cm² (mean over
   replicates) and `voltage_she_v`, its half-cell conversion. The
   conversion is affine, so correlation metrics are identical on either
   column; `voltage_she_v` is the comparable one. `targets_co2rr.csv`:
   one row per synthesis target with Faradaic efficiencies and production
   rates (mol cm⁻² s⁻¹) of H₂, CO, CH₄, C₂H₄ and liquid at a full-cell
   voltage of 3.3 V, by log-linear interpolation across the five current
   densities, no extrapolation. On the 2025-09-23 data this gives 280 HER
   rows and 279 CO₂RR rows over 52 element systems. Record the SHA-256 of
   both files in your result document.

2. **Folds** are leave-one-composition-out (LOCO) by *element system*: the
   `elements` column (`AgPd`, `AuCuPd`, …). Every ratio of the same set of
   elements is held out together, so a model never sees any measurement of
   the element system it is predicting. Leave-one-out (LOO) may be reported
   in addition, never instead.

3. **The metric** is `pearson_r2` from `evaluate.py`: the squared Pearson
   correlation of pooled out-of-fold predictions against the canonical
   targets. It is what the OCx24 paper reports as R², so it compares with
   the published baselines. `evaluate.py` also reports the coefficient of
   determination, MAE and Spearman ρ; report them too.

   ```
   python3 evaluate.py --targets targets/targets_her.csv --column voltage_she_v \
       --predictions predictions/her_loco.csv --scheme loco
   ```

   A predictions file has the columns `sample_id`, `prediction`, `fold`.
   `evaluate.py` reads the target values from the targets file, checks
   coverage and that each `fold` is the row's element system, and exits 1
   when a check fails. Reviewers re-score your tables with it.

## Rules that keep the protocol honest

- **Allowed inputs to a model:** the bound compositions (target and XRF),
  `source` and batch, any public computational data (the OCx24 descriptor
  file, OC20/OC22/OC25/OMat24 models, Materials Project, OQMD, Alexandria,
  the XRD phase identifications), any simulation you run yourself, and the
  published literature *about materials*, not about these samples.
- **Not allowed:** any electrochemical outcome of a held-out element system
  (voltage, Faradaic efficiency, production rate, or anything derived from
  them), from any file, including `dashboard`, the paper's processed tables,
  the dashboard's JSON exports, or your own earlier runs. Hyperparameters
  are fixed a priori or chosen inside the training folds. State in
  `leakage_audit` how you checked.
- **Do not redefine the task** to make it easier: no dropping of "hard"
  element systems, no replacing LOCO by a random split, no new target
  column. If you believe a different target is more physical, say so in a
  Note and report it alongside the fixed one.
- **Reproducibility:** the code archive re-runs from the bound inputs plus
  the external downloads you list with their hashes, and regenerates every
  predictions table the document names. Pin versions and seeds.

## Working on a Problem

After `start_attempt`, read the other Attempts' Notes; if a line of work is
well in hand, build on it (`builds_on`) or end your Attempt as abandoned with
a Note saying why. Write a Note when you pick features or a model family,
when a run beats or fails to beat the baseline, when you find a data issue,
and when you finish a phase; name the scripts you ran and what they showed.
Keep raw run logs in the code archive, not in Notes.

Cite the data as the OCx24 paper (arXiv:2411.11783) and the OCx25 dashboard
(CC-BY-4.0). The fairchem code is MIT.

## Independent agent-review pilot

Read `review.md` for the public review procedure and claim boundaries. Reviews
are supplied by community participants using the model specified in the pinned
protocol. No review, endorsement, private data or experimental work by the
original research team is assumed.

Pearson correlation squared is retained for comparison with the paper, but is
not sufficient for model acceptance. `evaluate.py` also returns signed
`pearson_r`. Model Problems require positive correlation, lower MAE and higher
coefficient-of-determination `r2` than the in-harness baseline, in addition to
their stated numerical bars. Report finite scores from complete canonical
prediction tables. A successful scorer run checks tables, not how training
was performed: reviewers must inspect and rerun the training procedure.

Measured XRF and XRD inputs support prediction after characterisation. A
proposal for an unmade catalyst must explain how each feature would be
available before synthesis, or use a separately evaluated model restricted to
such inputs. Do not substitute an assumed measured phase or composition
without identifying and evaluating that assumption.