cobra9786/engineering-log-triage-dataset

Engineering Log Triage Dataset

Summary

This dataset contains synthetic/sanitized engineering-log examples for structured fault triage.

Each example is formatted as a chat-style supervised fine-tuning record. The model input is an unstructured engineering report. The target assistant message is a strict JSON object containing a structured triage result.

This dataset was created for the LoRA-Adapted Engineering Log Triage Service project.

Intended Use

The dataset is intended for small-scale supervised fine-tuning and evaluation of instruction models on engineering fault-report triage.

The expected model behavior is to convert an unstructured report into validated JSON with these fields:

• category
• severity
• subsystem
• symptoms
• suspected_cause
• recommended_action
• requires_human_review

Dataset Structure

Files:

File	Split	Records
train.jsonl	train	48
validation.jsonl	validation	7
test.jsonl	test	3
dataset_manifest.json	metadata	n/a

Total records: 58

Each JSONL row has this shape:

{
  "id": "ENG-0001",
  "messages": [
    {
      "role": "system",
      "content": "..."
    },
    {
      "role": "user",
      "content": "Engineering report:\n\n..."
    },
    {
      "role": "assistant",
      "content": "{\"category\":\"sensor_or_signal_path\",...}"
    }
  ]
}

Label Schema

Categories

Allowed category values:

• sensor_or_signal_path
• communications
• power_or_battery
• firmware_or_software
• calibration_or_configuration
• mechanical_or_environmental
• data_pipeline_or_api
• unknown

Severity

Allowed severity values:

• low
• medium
• high
• critical

Human Review

requires_human_review is a boolean field indicating whether the generated triage result should be escalated for human inspection.

Data Creation

The examples are synthetic/sanitized and designed to mimic engineering, robotics, sensor, telemetry, firmware, configuration, mechanical, environmental, and data-pipeline fault reports.

The dataset was curated to support:

• prompt-only baseline evaluation
• LoRA/QLoRA supervised fine-tuning
• strict JSON/schema validation
• train/validation/test split discipline
• evaluation of output contract adherence

Split Policy

Current dataset version: 0.3.0

Split policy: lora_v1

The validation split was used for adapter selection. The test split was held out until after the selected adapter was chosen.

Limitations

This is a compact portfolio-scale dataset.

Known limitations:

• small total record count
• synthetic/sanitized examples
• small held-out test split
• not representative of all engineering domains
• not production validated
• severity classification needs more examples
• human-review policy needs more examples

The dataset should not be used to claim production-grade fault triage accuracy.

Related Project

This dataset supports a LoRA-adapted engineering-log triage service using:

• Qwen/Qwen2.5-1.5B-Instruct
• Hugging Face Transformers
• Datasets
• PEFT/LoRA
• TRL SFTTrainer
• bitsandbytes 4-bit quantization
• FastAPI
• Docker GPU serving

Evaluation Context

The dataset was used to compare a prompt-only Qwen baseline against a LoRA-adapted model.

The selected LoRA adapter improved strict JSON/schema compliance on validation and produced strict schema-valid JSON on the held-out test split.

Because the dataset is small, these results should be treated as directional evidence for a portfolio project rather than production accuracy estimates.

Citation

No formal citation is available. For portfolio or demonstration use, cite this dataset by its Hugging Face repository name.