PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools
Paper • 2604.01532 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
PF-D13
The agentic-scenario layer of PHMForge, reformatted into the unified SFT schema. Each scenario becomes a T-E3 (agentic tool-use) record; Cost-Benefit and Safety/Policy scenarios additionally become a T-D1 (decision) record.
The repository name is an internal code. See Provenance below for the underlying dataset.
Records
76 records. query = scenario question, annot = ground truth (answer + acceptance criteria, with the rationale removed). reasoning is null (these scenarios ship no chain-of-thought); the scenario's native templated rationale is preserved in a dedicated rationale field — a D13-specific column beyond the unified 7-field schema. 8 scenarios built on Week2-overlap datasets (CWRU / IMS / XJTU) are excluded (cross-paper de-dup) → 67 scenarios → 76 records.
Unified SFT schema (8 fields)
| field | type | meaning |
|---|---|---|
query |
str | the question / query / instruction |
image |
Image | null | always null in this dataset |
annot |
str | list[str] | ground-truth answer + acceptance criteria (the scenario rationale is removed and kept separately in rationale) |
reasoning |
str | null | always null here — these scenarios carry no chain-of-thought / thinking trace |
rationale |
str | null | D13-specific field. The scenario's native answer-justification — a templated one-liner derived from the ground-truth labels (e.g. “Based on ground truth RUL values from RUL_FD001.txt …”), not an LLM/CoT reasoning trace; null when the source has none |
cate |
"A".."E" | one of the five SFT categories (this dataset: E, D) |
task |
"T-xx" | unified task id (this dataset: T-E3 + T-D1) |
metadata |
str (JSON) | all other info; carries a "split" key when the source has train/val/test |
Load
from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/PF-D13")
Gated — request access on the dataset page; access is granted manually by the maintainers.
Provenance & license
This dataset is a reformatted derivative (unified SFT schema) of:
PHMForge — Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools (Columbia + IBM).
- Paper: https://arxiv.org/abs/2604.01532
- Code: https://github.com/DeveloperMindset123/PHMForge-A-Scenario-Driven-Agentic-Benchmark-for-Industrial-Asset-Lifecycle-Maintenance
Refer to the upstream source for the original licensing terms; this reformatted version is shared for research use. Please cite the upstream work.
Not yet included
Raw signal layer (T-C1 / T-C2) — not yet included. PHMForge's underlying sensor-signal datasets (13 PDMBench subsets + C-MAPSS + EngineMT-QA) map to time-series tasks whose input is signal → time-series image. That encoding step is owned by the team and not yet frozen, so the signal layer is intentionally not converted here yet; it will be added once the encoding is fixed.
- Downloads last month
- 9