smd20/socialengineering is a domain-specialized conversational model for social engineering, phishing awareness, and red-team education, fine-tuned from Google Gemma 4 E4B using Unsloth and exported as BF16 GGUF for efficient local deployment with llama.cpp, Ollama, LM Studio, and related runtimes.

The model was trained on a large bilingual Q&A corpus derived from authoritative social-engineering reference books, covering definitions, attack techniques (phishing, vishing, pretexting, baiting, tailgating), case studies, and defensive strategies.

Model Summary

Property	Value
Base architecture	Gemma 4 (E4B instruction-tuned variant)
Parameters	~8B
Precision / format	BF16 GGUF
Primary weight file	`unsloth-gemma-4-E4B-it.BF16.gguf`
Multimodal projector	`unsloth-gemma-4-E4B-it.BF16-mmproj.gguf`
Fine-tuning framework	Unsloth
Domain	Social engineering, phishing, red-team awareness
Languages	English, Persian (Farsi)
Context length (training)	2,048 tokens
Repository	smd20/socialengineering

Intended Use

Primary use cases

Organizational security-awareness chatbots
Phishing and social-engineering education for analysts and end users
Red-team / blue-team training scenarios in controlled environments
Local, privacy-preserving Q&A over social-engineering concepts

Out-of-scope / misuse

This model is not a substitute for legal, operational, or incident-response authority. It must not be used to conduct unauthorized attacks, harvest credentials, or deceive individuals outside approved training and research contexts.

Training Procedure

Fine-tuning was performed in Unsloth Studio on top of gemma-4-E4B, using a bilingual social-engineering Q&A corpus built from structured knowledge articles extracted from eight reference books.

Training hyperparameters

Setting	Value
Epochs	30
Learning rate	`2.0e-4`
Context length	2,048
LoRA rank	16
LoRA dropout	0.16
LoRA target modules	All enabled (`Enable LoRA`)
Optimizer	AdamW 8-bit
LR scheduler	Linear
Weight decay	0.001

Export configuration

Setting	Value
Training run	`gemma-4-E4B`
Export method	GGUF (quantized export path)
Published precision	BF16
Main artifact	`unsloth-gemma-4-E4B-it.BF16.gguf`

The published checkpoint preserves the merged fine-tuned weights in GGUF form for deployment with llama.cpp-compatible runtimes.

Training Data

The model was trained on conversational Q&A pairs grounded in curated social-engineering knowledge. The underlying datasets are publicly released on Hugging Face:

Dataset	URL	Records
English Q&A	https://huggingface.co/datasets/smd20/social-engineering-qa-english	3,330
Persian Q&A	https://huggingface.co/datasets/smd20/social-engineering-qa-persian	3,330

Reference corpora

Knowledge articles were derived from the following legally acquired books:

Deep Insight into Social Engineering
ESET Social Engineering Handbook
Learn Social Engineering: Learn the Art of Human Hacking (Erdal Ozkaya)
Social Engineering: How Crowdmasters, Phreaks, Hackers (Gehl & Lawson)
Social Engineering in Cybersecurity: Threats and Defenses (Gururaj et al.)
Social Engineering: The Science of Human Hacking (Christopher Hadnagy)
Social Engineering: The Art of Human Hacking (Christopher Hadnagy)
Sefreta: Zero to Hundred Social Engineering (Persian)

Corpus construction pipeline

Controlled segmentation of reference books
Schema-driven knowledge article generation (JSONL)
Grounded bilingual Q&A generation with strict source constraints
Global deduplication and bilingual split

Training Corpus Overview

    | Metric | Value |
    | --- | ---: |
    | English Q&A records | 3,330 |
    | Persian Q&A records | 3,330 |
    | Bilingual question units | 3,330 |
    | Total bilingual records (EN + FA) | 6,660 |
    | Structured knowledge articles | 1,165 |
    | Article coverage | 1,163 / 1,165 (99.8%) |
    | Reference books | 8 |
    | Deduplicated v1 duplicates skipped | 159 |

    ### Character-Length Statistics

    | Split | Field | Mean | Median | Std. Dev. | Min | Max |
    | --- | --- | ---: | ---: | ---: | ---: | ---: |
    | English | Question | 96.56 | 95.0 | 21.98 | 23 | 199 |
    | English | Answer | 180.12 | 171.0 | 80.13 | 3 | 827 |
    | Persian | Question | 81.08 | 80.0 | 21.76 | 12 | 181 |
    | Persian | Answer | 163.48 | 153.0 | 74.06 | 3 | 481 |
    | Combined (EN+FA) | Question | 88.82 | 88.0 | 23.2 | 12 | 199 |
    | Combined (EN+FA) | Answer | 171.8 | 161.0 | 77.6 | 3 | 827 |

    ### Knowledge Articles per Reference Book

    | Reference Book (internal ID) | Knowledge Articles |
    | --- | ---: |
    | Learn-Social-Engineering-Learn-the-Art-of-Human-Hacking-Dr.-Erdal-Ozkaya-_-WeLib.org-__FULL | 397 |

| Social-Engineering-Science-Hacking-Hadnagy_FULL | 239 | | Social-Engineering-Cybersecurity-Gururaj_FULL | 212 | | Social-Engineering-Crowdmasters-Gehl-Lawson_FULL | 206 | | Sefreta-Social-Engineering_FULL | 55 | | ESET-Social_engineering_handbook_FULL | 28 | | Social-Engineering-Art-Hacking-Hadnagy_FULL | 21 | | deep-insight-into-social-engineering_FULL | 7 |

Evaluation & Limitations

The model inherits base-model limitations and may hallucinate on out-of-domain queries.
Training data were LLM-assisted and should be complemented with human review for high-stakes deployments.
Copyright of source books remains with publishers; released datasets contain derived annotations only.
BF16 GGUF requires approximately 15.1 GB VRAM/RAM for full-precision loading.

How to Download from Hugging Face

Option 1 — `huggingface_hub` (recommended)

python

from huggingface_hub import hf_hub_download

repo_id = "smd20/socialengineering"
token = None  # set HF_TOKEN if the repo is private

model_path = hf_hub_download(
    repo_id=repo_id,
    filename="unsloth-gemma-4-E4B-it.BF16.gguf",
    token=token,
)
mmproj_path = hf_hub_download(
    repo_id=repo_id,
    filename="unsloth-gemma-4-E4B-it.BF16-mmproj.gguf",
    token=token,
)

print("Model:", model_path)
print("MMProj:", mmproj_path)

Option 2 — Snapshot download

python

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="smd20/socialengineering",
    allow_patterns=["*.gguf"],
)
print("Downloaded to:", local_dir)

Option 3 — CLI

bash

huggingface-cli download smd20/socialengineering \
  unsloth-gemma-4-E4B-it.BF16.gguf \
  unsloth-gemma-4-E4B-it.BF16-mmproj.gguf

Inference Examples

`llama.cpp`

llama-cli -hf smd20/socialengineering:BF16 --jinja

For multimodal usage:

llama-mtmd-cli -hf smd20/socialengineering:BF16 --jinja

`llama-cpp-python`

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="smd20/socialengineering",
    filename="unsloth-gemma-4-E4B-it.BF16-mmproj.gguf",
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "user",
            "content": "What is pretexting in social engineering, and how does it differ from impersonation?",
        }
    ],
)
print(response["choices"][0]["message"]["content"])

Ollama

ollama run hf.co/smd20/socialengineering:BF16

Authorship, Ownership, and Legal Notice

Legal owner and maintainer: Samad Sohrab — PhD Student in Artificial Intelligence.

This model checkpoint, its associated training configuration, and the derived Q&A datasets released under the smd20 Hugging Face namespace are authored and maintained by Samad Sohrab. All rights in the model card, training pipeline documentation, and derived dataset annotations are reserved by the author unless otherwise stated in the repository license.

Source-book copyrights remain with their respective publishers. This repository distributes fine-tuned model weights and derived instructional annotations only.

Acknowledgments

This work was conducted under the research supervision of Dr. Amir Nezami Safa, who served as academic advisor throughout dataset construction, model fine-tuning, and publication. His guidance on methodology, reproducibility, and scientific rigor was instrumental to this release.

Training infrastructure used Unsloth for efficient Gemma 4 fine-tuning and GGUF export.

Citation

If you use this model or the associated datasets in academic work, please cite:

bibtex

@misc{sohrab2026socialengineering,
  author       = {Sohrab, Samad and Nazami Saffa, Amir},
  title        = {Social Engineering Specialist: Fine-Tuned Gemma 4 E4B (GGUF)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/smd20/socialengineering}},
  note         = {PhD research release. Advisor: Dr. Amir Nazami Saffa}
}

Dataset Citations

bibtex

@misc{sohrab2026seqaen,
  author       = {Sohrab, Samad},
  title        = {Social Engineering Q&A Dataset (English)},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/smd20/social-engineering-qa-english}}
}

@misc{sohrab2026seqafa,
  author       = {Sohrab, Samad},
  title        = {Social Engineering Q&A Dataset (Persian)},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/smd20/social-engineering-qa-persian}}
}

Model card last updated: 2026-06-21T12:56:17.859588+00:00

smd20/socialengineering

Social Engineering Specialist — Gemma 4 E4B (GGUF)

Model Summary

Intended Use

Primary use cases

Out-of-scope / misuse

Training Procedure

Training hyperparameters

Export configuration

Training Data

Reference corpora

Corpus construction pipeline

Training Corpus Overview

Evaluation & Limitations

How to Download from Hugging Face

Option 1 — huggingface_hub (recommended)

Option 2 — Snapshot download

Option 3 — CLI

Inference Examples

llama.cpp

llama-cpp-python

Ollama

Authorship, Ownership, and Legal Notice

Acknowledgments

Citation

Dataset Citations

smd20/
socialengineering

Option 1 — `huggingface_hub` (recommended)

`llama.cpp`

`llama-cpp-python`