DestinyOtto/astris-minilm-l6-assetops-reranker

ASTRIS MiniLM-L6 AssetOps Reranker

This model is a fine-tuned cross-encoder reranker based on cross-encoder/ms-marco-MiniLM-L6-v2.

It was fine-tuned for industrial maintenance and asset operations ranking tasks using selected clean examples from ibm-research/AssetOpsBench.

Intended Use

This reranker scores (query, candidate) pairs for maintenance retrieval workflows. It can be used after initial embedding retrieval to rerank candidate passages, sensors, failure modes, diagnostic hypotheses, work-order actions, or maintenance knowledge chunks.

In ASTRIS, this model is intended to improve retrieval quality for industrial maintenance reasoning, predictive maintenance support, failure-mode analysis, sensor-to-fault mapping, and work-order recommendation workflows.

Training Data

Selected clean AssetOpsBench files were used:

• failuresensoriq_standard/all.jsonl
• failuresensoriq_standard/all_multi_answers.jsonl
• task/failure_mapping_senarios.jsonl
• task/phm_utterance.jsonl
• task/rule_monitoring_scenarios.jsonl
• asset/compressor_utterance.jsonl
• asset/hydrolicpump_utterance.jsonl

The perturbed AssetOpsBench files were not used for training in this run.

Base Model

cross-encoder/ms-marco-MiniLM-L6-v2

Model Size

The main trained weight file is approximately 86.66 MB.

Evaluation

The repository includes corrected reranker metrics comparing the raw base model and the fine-tuned model on the same held-out AssetOpsBench test groups.

Metrics include:

• Top-1 accuracy
• Exact match@k
• Precision@k
• Recall@k
• F1@k
• MRR
• nDCG@k

The corrected evaluation compares two separately loaded models:

1. The raw base model: cross-encoder/ms-marco-MiniLM-L6-v2
2. The fine-tuned ASTRIS reranker

This avoids accidentally evaluating the same trained model twice.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "Desimulator/astris-minilm-l6-assetops-reranker"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

query = "For compressor, what are the key failure modes when speed has abnormal readings?"

candidates = [
    "valve fault",
    "cooling system fault",
    "compressor stall",
    "damaged impeller",
    "misalignment"
]

inputs = tokenizer(
    [query] * len(candidates),
    candidates,
    padding=True,
    truncation=True,
    return_tensors="pt"
)

with torch.no_grad():
    scores = torch.sigmoid(model(**inputs).logits.squeeze(-1))

ranked = sorted(
    zip(candidates, scores.tolist()),
    key=lambda x: x[1],
    reverse=True
)

print(ranked)

Dataset Citation

This model was trained using selected clean examples from ibm-research/AssetOpsBench.

If you use this model or the dataset-derived training setup, please cite the original AssetOpsBench work:

@misc{patel2025assetopsbenchbenchmarkingaiagents,
  title={AssetOpsBench: Benchmarking AI Agents for Task Automation in Industrial Asset Operations and Maintenance},
  author={Dhaval Patel and Shuxin Lin and James Rayfield and Nianjun Zhou and Roman Vaculin and Natalia Martinez and Fearghal O'donncha and Jayant Kalagnanam},
  year={2025},
  eprint={2506.03828},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2506.03828}
}

Limitations

This model was fine-tuned on a compact industrial maintenance benchmark, not a full proprietary plant history. It should be used as a reranking component, not as the only source of truth for safety-critical maintenance decisions.

The model scores relevance between a query and candidate text. It does not directly calculate RUL, inspect raw sensor time series, or replace certified maintenance judgment.