Mohamud24/gemma-4-tech-interviewer

Gemma 4 E2B - Technical Interviewer (Bilingual: English & Somali)

A fine-tuned version of google/gemma-4-e2b-it, trained to conduct structured technical job interviews and mock training sessions in both English and Somali (af Soomaali).

Model Description

This model was fine-tuned using QLoRA (4-bit quantization + LoRA) on a custom bilingual dataset of 7,860 structured interview conversations spanning 10 technology domains.

The model can:

• Ask domain-specific technical interview questions (Frontend, Backend, DevOps, Data Science, Cybersecurity, etc.)
• Score candidate answers on a 0-10 scale with structured JSON output
• Open and close both Mock Training Sessions (with coaching feedback) and Company Hiring Sessions (neutral/professional)
• Operate fully in Somali or English depending on the session language

Training & Verification Details

• Base Model: google/gemma-4-e2b-it
• Fine-tuning Method: QLoRA (4-bit NF4)
• LoRA Rank (r): 16
• LoRA Alpha: 32
• Training Steps: 1,572
• Dataset Size: 7,860 rows
• Total Tokens Evaluated: 1,573,545 tokens
• Hardware: Google Colab Pro (A100 GPU)

Empirical Evaluation Metrics

The proposed fine-tuned model was evaluated on a held-out validation dataset:

• Token Accuracy (Correct Predictions): 82.48%
• Prediction Error Rate: 17.52%
• Final Training Loss: 0.6198
• Final Validation Loss: 0.7648
• Validation Perplexity: 2.15 (down from 288.40 on base model)

Supported Tasks

• ask_technical_question: Ask a domain-specific interview question
• score_candidate_answer: Score an answer with JSON output (0-10)
• open_mock_interview_session: Open a bilingual mock/training session
• close_mock_interview_session: Close with coaching feedback and scores
• open_hiring_interview_session: Open a formal company hiring session
• close_hiring_interview_session: Close hiring session professionally

Languages

• en - English
• so - Somali (af Soomaali)

Citation

@misc{gemma4-technical-interviewer-2026, title={Fine-tuned Gemma 4 E2B for Bilingual Technical Interviewing}, author={Mohamud Abdikarim}, year={2026}, publisher={Hugging Face}, url={https://huggingface.co/Mohamud24/gemma-4-e2b-technical-interviewer} }