README.md

---
license: cc-by-4.0
task_categories:
- time-series-forecasting
- other
tags:
- system-identification
- control-theory
- dynamics
- synthetic
- benchmark
- nonlinear-dynamics
size_categories:
- 100K<n<1M
pretty_name: PLANTFORGE
---

# PLANTFORGE corpus

A procedural control-plant corpus for in-context system identification,
organized along three axes jointly: **nonlinearity family × excitation class
× sampling rate (exact ZOH)**, with per-instance **Fisher-information
identifiability annotations**.

Code, generator, training/evaluation pipeline, and full documentation:
https://github.com/Soarr01/plantforge (see `docs/DATASHEET.md` there for the
full Datasheet-for-Datasets writeup this card summarizes).

## Quick facts

- **240,000 instances**: 5 families × 4 excitations × 3 rates × 4000
  instances/cell.
- **Families**: `stribeck` (velocity friction) · `backlash` (input deadzone)
  · `saturate` (input clipping) · `boucwen` (output hysteresis) ·
  `drivetrain` (two-inertia motor/gear/compliant-load).
- **Excitations**: `prbs` · `multisine` · `chirp` · `closedloop` (true
  sequential PI loop, not a two-pass imitation).
- **Rates**: 10 / 20 / 50 Hz, exact zero-order-hold from one shared
  continuous-time truth per instance (state-nonlinear families substep
  internally so dt and dt/4 agree at common instants).
- **Ground truth**: named physical parameters per instance, no hidden
  normalization (verified: re-simulation from θ reproduces y to 1e-6).
- **Identifiability annotations**: per-parameter relative Cramér-Rao lower
  bound and FIM log10-condition-number, per (instance, excitation, rate).
  Note an honest caveat from the release experiments: these annotations do
  **not** positively predict in-context *prediction* difficulty (within-cell
  median Spearman r ≈ −0.09 against per-instance prediction nMSE, robust to
  confound controls) — parameter-recovery difficulty decouples from
  prediction difficulty. They are intended as metadata for excitation-design
  and identifiability studies, not as a difficulty score.

## Format

60 shard files (`{family}_{excitation}_dt{rate}hz.pt`, PyTorch tensor dicts),
plus `registry.json`. Each shard:

```python
import torch
shard = torch.load("stribeck_multisine_dt50hz.pt")
shard["u"], shard["y"]          # (T, B) input/output trajectories
shard["theta"]                  # (B, K) named physical parameters
shard["keys"]                   # length-K parameter names
shard["rel_crlb"]               # (B, K) per-parameter relative CRLB
shard["log10_cond"]             # (B,) log10 FIM condition number
shard["dt"], shard["family"], shard["excitation"]   # shard metadata
```

## Updates

**2026-07-27**: corrected the `rel_crlb`/`log10_cond` identifiability
annotations. The generator's `identifiability()` function previously used
`torch.linalg.pinv` on the regularized Fisher information matrix, which
silently returned `rel_crlb=0` for any parameter with near-zero sensitivity
to the excitation (e.g. `sat` when the excitation never reaches the
saturation limit) instead of a large value — the opposite of correct CRLB
semantics (zero sensitivity means maximally *unidentifiable*, which should
show a *large* rel_crlb, not a near-zero one). This affected 9.3% of all
240,000 rows (46.4% of the `saturate` family). The annotations were
recomputed in place with a corrected regularized inverse; `u`, `y`, and
`theta` are byte-identical to the previous revision — only `rel_crlb` and
`log10_cond` changed. The previous revision remains addressable via its
commit SHA on the Files and versions tab.

## License

CC BY 4.0 — see `LICENSE-DATA` in the code repository, or
https://creativecommons.org/licenses/by/4.0/. Attribution: cite the code
repository (https://github.com/Soarr01/plantforge) and this dataset.

## Citation

`<BibTeX — TBD, pending paper draft>`