AEGIS Bilingual Industrial Action Classifier is an English-Arabic controlled industrial-action classification prototype built on bigscience/mt0-small using LoRA/PEFT.

It was developed as part of AEGIS AI, an industrial intelligence portfolio platform for manufacturing, robotics, warehouses, logistics, and related operational environments.

Instead of generating unrestricted operational instructions, the model selects one of eight predefined action classes from operational evidence and a question.

The selected class is then mapped deterministically to an approved English or Arabic response.

Key Features

English and Arabic industrial classification
bigscience/mt0-small base model
LoRA / PEFT fine-tuning
Eight controlled industrial action classes
Single-token numeric classification targets
Candidate scoring across all eight classes
Deterministic approved-response rendering
Free-form operational action generation disabled
FastAPI backend integration
React frontend integration
CPU-trained portfolio prototype

Architecture

The AEGIS inference architecture is:

text

Operational Evidence
        +
Industrial Question
        +
Language (EN / AR)
        |
        v
bigscience/mt0-small
        +
AEGIS LoRA Adapter
        |
        v
Candidate scoring over
single-token codes 0-7
        |
        v
Selected Class Code
        |
        v
Semantic Action ID
        |
        v
Deterministic Approved
English / Arabic Response

The model therefore functions as a controlled classifier, not as an unrestricted industrial chatbot.

Free-form operational action generation is disabled.

Controlled Action Classes

Code	Semantic Action ID	Industrial Module
`0`	`ALERT_ESCALATION`	AI Alerts
`1`	`DOCUMENT_GROUNDED_ANSWER`	Document Assistant
`2`	`FACTORY_DEGRADED_REVIEW`	Factory Status
`3`	`PREDICTIVE_MAINTENANCE`	Predictive Maintenance
`4`	`ROBOT_SAFE_RESTART`	Robot Monitoring
`5`	`VISION_QUARANTINE`	Vision Inspection
`6`	`WORKER_SAFETY_STOP`	Worker Safety
`7`	`WORK_ORDER_CREATE`	Workflow Automation

The complete English and Arabic mappings are provided in:

class_code_map.json

Approved Actions

0 — ALERT_ESCALATION

English

Acknowledge, assess severity, and escalate according to the alert playbook.

Arabic

أكد استلام التنبيه، وقيّم مستوى الخطورة، وصعّد الحالة وفق دليل إجراءات التنبيهات.

1 — DOCUMENT_GROUNDED_ANSWER

English

Answer only from approved evidence and cite the source.

Arabic

أجب فقط استنادًا إلى الأدلة المعتمدة، واذكر المصدر.

2 — FACTORY_DEGRADED_REVIEW

English

Review line KPIs and investigate the assets driving the degraded status.

Arabic

راجع مؤشرات الأداء الرئيسية (KPIs) لخط الإنتاج، وتحقق من الأصول المتسببة في تدهور الحالة.

3 — PREDICTIVE_MAINTENANCE

English

Inspect the asset, confirm the trend, and schedule maintenance before failure.

Arabic

افحص الأصل، وتأكد من اتجاه التدهور، وجدول أعمال الصيانة قبل حدوث العطل.

4 — ROBOT_SAFE_RESTART

English

Diagnose the protective stop and verify safe conditions before restart.

Arabic

شخّص سبب التوقف الوقائي، وتحقق من توفر ظروف آمنة قبل إعادة التشغيل.

5 — VISION_QUARANTINE

English

Quarantine the suspect item and send it for secondary inspection.

Arabic

اعزل العنصر المشتبه به وأرسله لإجراء فحص ثانوي.

6 — WORKER_SAFETY_STOP

English

Stop unsafe work, secure the area, and follow the approved safety procedure.

Arabic

أوقف العمل غير الآمن، وأمّن المنطقة، واتبع إجراء السلامة المعتمد.

7 — WORK_ORDER_CREATE

English

Create and track a work order with owner, priority, evidence, and closure verification.

Arabic

أنشئ أمر عمل وتابعه مع تحديد المسؤول، والأولوية، والأدلة، والتحقق من الإغلاق.

Dataset

Training dataset:

syed7741/aegis-bilingual-industrial-ai-dataset

The dataset contains 64 unique synthetic industrial records across eight operational modules.

Languages:

English
Arabic

The final experiment used:

Dataset Property	Value
Unique records	64
Training records	48
Held-out records	16
Training records per module	6
Held-out records per module	2
Expanded training examples	192
Bilingual evaluation examples	32
Approved action classes	8

The record-level train/evaluation split was performed before prompt expansion.

The dataset design also includes:

evidence-aligned examples
module labels removed from questions
target-leakage checks
balanced module representation
bilingual English-Arabic evidence

Training Configuration

Setting	Value
Base model	`bigscience/mt0-small`
Architecture	MT0 + LoRA
Epochs	10
Learning rate	`5e-4`
Batch size	1
Gradient accumulation steps	4
Maximum input length	512
LoRA rank	8
LoRA alpha	16
LoRA dropout	0.05
LoRA target modules	`q`, `v`
Trainable parameters	344,064
Total parameters	300,520,832
Trainable percentage	~0.1145%
Training hardware	CPU
PyTorch	2.13.0+cpu
Transformers	5.15.0
PEFT	0.19.1

Initial evaluation loss:

2.760820

Final evaluation loss:

0.004219

Total recorded training time:

1004.12 seconds

Approximately:

16.7 minutes

Held-Out Evaluation

The final model was evaluated using a balanced set of 16 held-out records.

Every held-out record was tested in both English and Arabic.

This produced:

32 classification tests

Results

Language	Correct	Total	Accuracy
English	16	16	100%
Arabic	16	16	100%
Combined	32	32	100%

Each of the eight industrial modules contained two held-out records.

For every module, the classifier achieved:

text

English: 2 / 2
Arabic:  2 / 2
Combined: 4 / 4

The detailed benchmark results are available in:

controlled_action_evaluation.json

Important Benchmark Interpretation

The 100% accuracy result applies only to this specific small, balanced, synthetic held-out benchmark.

It does not mean that the model has 100% accuracy in:

real factories
refineries
warehouses
logistics facilities
robotics deployments
real worker-safety incidents
real predictive-maintenance environments
other unseen industrial conditions

No real-world industrial safety certification, regulatory validation, or field-validation claim is made.

The benchmark demonstrates successful controlled classification on the project's synthetic evaluation set.

Why Single-Token Numeric Classes?

Earlier AEGIS experiments used long semantic labels such as:

text

WORKER_SAFETY_STOP
PREDICTIVE_MAINTENANCE
DOCUMENT_GROUNDED_ANSWER

These labels tokenize into different sequence lengths.

For a sequence-to-sequence model, unequal label lengths can introduce undesirable generation and scoring behavior.

The final architecture therefore uses:

text

Each class code is represented as a single tokenizer token.

The numeric code is then mapped deterministically to the corresponding semantic industrial action.

This separates:

Model Classification

from:

Operational Response Rendering

Candidate Scoring

The AEGIS runtime does not rely on unrestricted text generation to decide the operational action.

Instead:

A bilingual prompt is constructed from evidence, question, language, and the approved action menu.
The prompt is passed through MT0 + the LoRA adapter.
All eight candidate codes are evaluated.
A loss value is calculated for every candidate.
The candidate with the lowest loss is selected.
The code is mapped to a semantic action ID.
The final response is loaded from the approved action catalog.

Conceptually:

text

Evidence + Question
        |
        v
Score class 0
Score class 1
Score class 2
Score class 3
Score class 4
Score class 5
Score class 6
Score class 7
        |
        v
Lowest Loss
        |
        v
Selected Action

Controlled Response Design

The final operational response itself is not generated freely by the model.

For example:

Predicted class: 6

maps to:

WORKER_SAFETY_STOP

which maps to the approved English response:

Stop unsafe work, secure the area, and follow the approved safety procedure.

or the approved Arabic response:

أوقف العمل غير الآمن، وأمّن المنطقة، واتبع إجراء السلامة المعتمد.

This architecture was chosen to demonstrate a controlled-output design for an industrial AI prototype.

Example — English

Evidence

text

An unsafe condition has been confirmed around CONV-02.
Personnel are still performing work near the hazard,
access to the affected area remains open, and required
safety controls have not yet been completed.

Question

text

Based on the operational evidence, what action should be
taken for CONV-02?

Selected class

6

Semantic action

WORKER_SAFETY_STOP

Approved response

Stop unsafe work, secure the area, and follow the approved safety procedure.

Example — Arabic

الأدلة التشغيلية

text

تم تأكيد وجود حالة غير آمنة بالقرب من المعدة CONV-02.
لا يزال بعض العاملين يواصلون العمل بالقرب من مصدر الخطر،
كما أن الوصول إلى المنطقة المتأثرة ما زال مفتوحًا ولم تكتمل
ضوابط السلامة المطلوبة بعد.

السؤال

استنادًا إلى الأدلة التشغيلية، ما الإجراء الذي ينبغي اتخاذه للمعدة CONV-02؟

الفئة المختارة

6

Semantic Action ID

WORKER_SAFETY_STOP

الإجراء المعتمد

أوقف العمل غير الآمن، وأمّن المنطقة، واتبع إجراء السلامة المعتمد.

Loading the LoRA Adapter

The repository contains a PEFT LoRA adapter.

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from peft import PeftModel

repo_id = "syed7741/aegis-bilingual-industrial-action-classifier"

tokenizer = AutoTokenizer.from_pretrained(repo_id)

base_model = AutoModelForSeq2SeqLM.from_pretrained(
    "bigscience/mt0-small"
)

model = PeftModel.from_pretrained(
    base_model,
    repo_id,
)

model.eval()

Loading the LoRA adapter alone does not reproduce the complete AEGIS controlled inference pipeline.

The full system additionally performs candidate scoring across all eight numeric codes and deterministic response lookup using:

class_code_map.json

AEGIS AI Application Integration

The classifier is integrated into the larger AEGIS AI application.

The application includes:

React frontend
TypeScript
Material UI
FastAPI backend
PostgreSQL
JWT authentication
role-based access control
local Hugging Face models
local RAG
industrial document assistance
robot monitoring
worker safety
predictive maintenance
vision inspection
workflow automation
AI alerts
industrial action classification

The Industrial Action AI interface supports:

English evidence
Arabic evidence
English questions
Arabic questions
controlled classification
approved response rendering
class-code display
semantic action IDs
decision-margin display

Score Margin

The AEGIS API may expose a value called:

score_margin

This represents the difference between the best candidate loss and the next-best candidate loss.

It is useful for prototype diagnostics.

It is not a calibrated probability.

It should therefore not be interpreted as:

95% confidence

or another probabilistic confidence measurement without separate calibration.

Intended Uses

This repository is intended for:

AI engineering demonstrations
bilingual NLP experimentation
industrial AI portfolio work
LoRA / PEFT demonstrations
controlled classification research
English-Arabic model experimentation
deterministic response-catalog architectures
educational demonstrations

Out-of-Scope Uses

The model should not be used as an autonomous decision maker for:

industrial machinery
emergency shutdown systems
worker safety systems
medical systems
life-critical infrastructure
autonomous robot control
regulatory compliance decisions

without appropriate real-world validation, engineering controls, qualified human oversight, and applicable certification.

Limitations

Important limitations include:

The dataset is synthetic.
Only 64 unique records were used.
Only eight action classes are supported.
Evaluation is performed on a small held-out synthetic benchmark.
The model has not been validated on large real-world industrial datasets.
The model may fail on out-of-distribution inputs.
The score margin is not calibrated confidence.
Real operational language may differ significantly from training examples.
Arabic dialect variation was not comprehensively evaluated.
Industrial terminology outside the dataset may be misunderstood.
The model does not replace engineering judgment.
The model has not undergone industrial safety certification.

Safety

AEGIS AI is a research and portfolio prototype.

It must not be treated as a substitute for:

qualified engineers
plant operators
maintenance professionals
safety officers
emergency-response teams
official operating procedures
manufacturer documentation
regulatory requirements
site-specific safety rules

Human review is required for real-world operational decisions.

The controlled-action architecture is intended to demonstrate a more constrained output pattern for an industrial AI prototype.

It does not make the system inherently safe or certified for autonomous industrial deployment.

Repository Files

File	Purpose
`adapter_config.json`	PEFT LoRA configuration
`adapter_model.safetensors`	Trained LoRA adapter weights
`tokenizer.json`	Tokenizer artifact
`tokenizer_config.json`	Tokenizer configuration
`class_code_map.json`	Numeric class → semantic action → EN/AR approved action
`controlled_action_evaluation.json`	Detailed held-out evaluation results
`training_summary.json`	Training configuration and experiment metadata
`README.md`	Model card

Related Resources

Dataset

syed7741/aegis-bilingual-industrial-ai-dataset

Base Model

bigscience/mt0-small

AEGIS AI GitHub Repository

syedasim7741/AEGIS-AI

Reproducibility

The AEGIS AI GitHub repository contains the training and evaluation scripts used during development, including:

text

build_aegis_bilingual_dataset.py

train_aegis_bilingual_lora_cpu.py

train_aegis_bilingual_mt0_lora_cpu.py

train_aegis_controlled_action_mt0_lora_cpu.py

train_aegis_controlled_action_balanced_mt0_lora_cpu.py

train_aegis_single_token_balanced_mt0_lora_cpu.py

train_aegis_single_token_balanced_mt0_lora_512_cpu.py

evaluate_aegis_bilingual_mt0.py

The final winning experiment is represented by:

train_aegis_single_token_balanced_mt0_lora_512_cpu.py

Author

Sayyad Asim

AEGIS AI
Industrial Intelligence Platform

Hugging Face:

syed7741

GitHub:

syedasim7741

Portfolio Note

AEGIS AI was developed as an AI-engineering portfolio project demonstrating the combination of:

bilingual model development
synthetic dataset design
LoRA fine-tuning
controlled classification
RAG
FastAPI
React
industrial AI architecture
model evaluation
safety-aware output constraints

The project is a prototype and does not claim production certification.