PHMForge

Canonical copy: ibm-research/PHMForge. This repository is a mirror; cite and download from the IBM Research copy.

PHMForge is a benchmark for LLM agents on industrial Prognostics and Health Management (PHM). It has 99 scenarios across eight asset classes. To solve one, an agent has to find and call MCP-served tools that wrap PHM algorithms, then return an answer that is checked against a deterministic ground truth.

• Paper: PHMForge: Evaluating LLM Agents on Industrial Prognostics through MCP-Native, Algorithm-Grounded Tools, NeurIPS 2026 Evaluations & Datasets Track (poster). Link to be added.
• Code: the MCP server implementations and evaluation harness are not part of this release.
• AssetOpsBench: PHMForge is a companion benchmark to AssetOpsBench. 45 of its scenarios also ship in the AssetOpsBench PHM release (see Relation to the AssetOpsBench PHM release).
• License: CC BY 4.0 for the scenario files. The underlying sensor datasets keep their own licenses (see Source datasets).

Harbor tasks

The harbor/ folder packages all 99 scenarios as 123 Harbor tasks: the 24 lithium-ion scenarios in two phrasings, plus the other 75. Each task carries the question, the public data it needs, and a separate verifier. See harbor/README.md.

Dataset structure

There is one test split. The configs are:

Config	Scenarios	Content
all (default)	99	everything
rotating_aero	75	turbofans, aero-engines, bearings, motors, gearboxes, industrial engine
lithium_ion	24	NASA PCoE lithium-ion cells (BESS extension)

Per task category: Health Analysis 39, RUL Prediction 20, Fault Classification 20, Safety/Policy Evaluation 13, Cost-Benefit Analysis 7.

Per asset class: aero-engine 30, lithium-ion cell 24, bearing 22, turbofan 14, electric motor 3, gearbox 3, induction motor 2, industrial engine 1.

Fields

Field	Type	Description
id	string	Scenario ID (e.g. pdm_rul_001, engine_health_combined_005, bess_rul_003)
subset	string	rotating_aero or lithium_ion
task_category	string	One of the five task categories
asset_class	string	Asset class
source_dataset	string	Public dataset the ground truth is computed on (e.g. CMAPSS_FD001, NASA_PCoE_B0005)
question	string	Scenario query. For lithium_ion this is the protocol-style ("explicit") form.
question_operator_style	string	lithium_ion only: the operator-style ("fuzzy") form of the same question
required_tools	list[string]	Tools in the reference workflow
procedure	string	Reference procedure
ground_truth	string (JSON)	Ground truth and validation tolerances (e.g. expected_mae_range, labels, required fields)
metadata	string (JSON)	Remaining original fields (e.g. dependency analysis, difficulty, failure modes)
known_issue	string	Empty, or a description of a documented problem with the scenario (see Known issues)

from datasets import load_dataset
ds = load_dataset("yushengli/PHMForge", "all", split="test")
import json; gt = json.loads(ds[0]["ground_truth"])

Datasheet

Motivation

LLM agents are starting to call industrial asset-management tools through the Model Context Protocol (MCP). PHMForge measures whether an agent can select, sequence and compose algorithm-grounded tools for PHM tasks, rather than whether it can produce a plausible text answer. It was created by the PHMForge team (Columbia University, IBM Research, Georgia Tech).

Composition

• What one item is: a scenario. It has a natural-language query, the reference tool workflow, and a ground truth with validation tolerances.
• Data volume: the scenarios point to public sensor datasets but do not include them.
• Personal information: none; there are no human subjects.
• Agent outputs: the dataset contains none.

Collection process

• 75 rotating-equipment and aero-engine scenarios: written by the consortium's subject-matter experts in six expansion stages (paper, Appendix D).
• 24 lithium-ion scenarios: added in Stage 6.
• Ground truth: derived from the source datasets with documented scripts, or cited to the source.
• LLM use: the paper states that no LLM was used in scenario generation, query formulation or ground-truth derivation.

Quality assessment

• Human rating: two SMEs independently rated a stratified sample of 30 of the 75 rotating-equipment/aero-engine scenarios on realism, difficulty calibration and ground-truth correctness (1–4 scale). Ordinal Krippendorff's α was 0.86, 0.58 and 0.73 respectively, and 0.75 pooled (paper, Table 2).

• The 24 lithium-ion scenarios are not covered by that sample. As a supplementary check:

  • the checkable ground truths were recomputed from the processed NASA PCoE data;
  • two LLM-originated rating columns were produced under the same rubric, then reviewed and revised by an author. The paper reports their post-review agreement (Table 2, Appendix E.6).

  These are not independent human ratings or an inter-annotator agreement between experts. The issues found are listed under Known issues.

Known issues

61 of the 99 scenarios carry a non-empty known_issue field. They fall into two groups.

Errors to be fixed in v1.1:

• Dataset named in the question text (10 RUL scenarios). In nine of them the question names a bearing dataset, while the ground truth is computed on the C-MAPSS subset in source_dataset. The field is correct and the question text is not. A tenth (pdm_rul_005) names IMS for FEMTO data.
• Contradictions (9 aero-engine Health scenarios). Scenarios with the same signal summary have different ground truths (013–017 and 018–020), and engine_health_combined_001 has an internally inconsistent ground truth.
• Lithium-ion ground truth or specification errors (9 scenarios). These were found by the recomputation check (paper, Appendix E.6). None of them is an RUL Prediction scenario.

Generic ground truth (33 scenarios):

• Fault Classification: a generic accuracy range plus a placeholder field.
• Cost-Benefit: cost ranges that depend on unstated unit costs.
• Safety/Policy: only the allowed risk levels are specified.
• FEMTO RUL: ranges without a unit.

These scenarios can be scored only on what the ground truth specifies.

Scoring: the results in the paper scored all scenarios as released.

See RELEASE_NOTES.md for the per-scenario list. Until v1.1, report results both with and without the flagged scenarios.

Relation to the AssetOpsBench PHM release

• The 45 RUL, Fault, Safety and Cost-Benefit scenarios are identical (same IDs and text) to the prognostics_and_health_management config of ibm-research/AssetOpsBench (code: github.com/IBM/AssetOpsBench).
• The 30 aero-engine Health scenarios differ.
  • Here they are the engine_health_combined_* formulations, which the paper's automated harness evaluated (Table 3).
  • AssetOpsBench lists the original EngineMTQA formulations (enginemtqa_train_qa_*).
  • Both are derived from EngineMTQA. The ID correspondence between the two formulations will be added in a later release.

Recommended uses

• Evaluating tool-using agents on PHM workflows: tool selection, ordering, parameterization, and answer correctness against the deterministic ground truth.
• Not recommended:
  • Treating a high score as evidence that an agent is ready for unsupervised safety-critical operation.
  • Drawing per-class conclusions for the four smallest asset classes (n ≤ 3 each).

Distribution and maintenance

• Distribution: through Hugging Face under CC BY 4.0.
• Versioning: versioned increments under the living-benchmark expansion plan (paper, Appendix D.3). Issues and corrections go through the code repository's issue tracker.

Source datasets

The scenarios are built on these public datasets. If you use them, cite the original sources; their licenses and terms of use apply to the sensor data, including the processed subsets in harbor/.

Dataset	Asset	Original reference	Accessed via
C-MAPSS FD001–FD004	Turbofan	Saxena et al., PHM 2008, doi:10.1109/PHM.2008.4711414	NASA Prognostics Data Repository
NASA PCoE Li-ion (B0005–B0018)	Li-ion cell	Saha & Goebel, Battery Data Set, NASA Ames, 2007	NASA Prognostics Data Repository
EngineMTQA	Aero-engine	Wang et al., ITFormer, ICML 2025, PMLR 267	Hugging Face
Azure Predictive Maintenance	Industrial engine	Microsoft Azure AI Gallery, Kaggle mirror	PDMBench
CWRU	Bearing	Case Western Reserve University Bearing Data Center	PDMBench
FEMTO / PRONOSTIA	Bearing	Nectoux et al., IEEE PHM 2012	PDMBench
IMS	Bearing	Qiu et al., J. Sound Vib. 2006, doi:10.1016/j.jsv.2005.03.007	PDMBench
XJTU-SY	Bearing	Wang et al., IEEE Trans. Reliab. 2020, doi:10.1109/TR.2018.2882682	PDMBench
HUST	Bearing	Thuan & Hong, BMC Res. Notes 2023, doi:10.1186/s13104-023-06400-4	PDMBench
MFPT	Bearing	Bechhoefer, MFPT fault data sets, 2013	PDMBench
Paderborn	Bearing	Lessmeier et al., PHME 2016, doi:10.36001/phme.2016.v3i1.1577	PDMBench
Rotor Broken Bar	Induction motor	Treml et al., IEEE DataPort 2020, doi:10.21227/fmnm-bn95	PDMBench
UoC Gearbox	Gearbox	Cao et al., IEEE Access 2018, doi:10.1109/ACCESS.2018.2837621	PDMBench
Mendeley bearing	Bearing	original source not confirmed	PDMBench
Electric Motor Vibrations	Electric motor	original source not confirmed	PDMBench
Planetary PdM	Planetary gearbox	original source not confirmed	PDMBench

PDMBench: Zhang et al., OpenReview; processed files in submission096 on Hugging Face (CC BY 4.0).

Citation

@inproceedings{li2026phmforge,
  title     = {{PHMForge}: Evaluating {LLM} Agents on Industrial Prognostics through
               {MCP}-Native, Algorithm-Grounded Tools},
  author    = {Li, Yusheng and Feng, Tianjun and Chen, Yunfeng and Tsai, Chun-Yi and Sun, Yihan and
               Patel, Dhaval and El Maghraoui, Kaoutar and Lin, Shuxin and Das, Ayan},
  booktitle = {Advances in Neural Information Processing Systems (Evaluations and Datasets Track)},
  year      = {2026}
}