---
license: mit
language:
- en
pipeline_tag: text-generation
tags:
- jobs
- HR
- Interview
- Question
- job
metrics:
- bertscore
base_model:
- openai-community/gpt2
library_name: transformers
---

# Model Card for EchoHire Question Generator (DistilGPT-2 Fine-tuned)

This model is a fine-tuned DistilGPT-2 designed to generate **interview questions** based on a job title and associated skills. It was trained on a curated dataset of ~100 jobs, each with 17–19 questions, to produce relevant, structured interview questions for recruiters or learning platforms.

## Model Details

### Model Description

This model generates questions automatically given a prompt with the job title and skills. The outputs are intended to help recruiters, HR teams, or training platforms quickly generate relevant interview questions without manually writing each one.

- **Developed by:** Syed Zeeshan Shah  
- **Model type:** Causal Language Model (GPT-2)  
- **Language(s) (NLP):** English  
- **License:** MIT  
- **Finetuned from model:** distilgpt2  
- **Shared by:** Zeeshan506  

### Model Sources

- **Repository:** [Hugging Face Model Hub](https://huggingface.co/Zeeshan506/echohire-qgen-distilgpt2)  
- **Demo:** N/A  

## Uses

### Direct Use

- Generate interview questions automatically by providing a job title and skills.  
- Can be used for recruitment platforms, HR automation, or interview prep content.

### Downstream Use

- Could be fine-tuned further on **domain-specific roles** (e.g., data science, embedded systems).  
- Can be integrated into apps, bots, or SaaS platforms for HR automation.

### Out-of-Scope Use

- Not intended for generating **biased or discriminatory questions**.  
- Should not be used as a sole source of assessment or evaluation for candidates.  
- Outputs may not always reflect up-to-date technology trends or best practices.

## Bias, Risks, and Limitations

- Trained on a **limited dataset of ~100 jobs**, so may not generalize to rare or niche roles.  
- Model can produce long or merged questions; post-processing may be needed.  
- Users should verify all generated questions for correctness and appropriateness.

### Recommendations

- Review outputs before using in real interviews.  
- Consider further fine-tuning for specific industries or technical domains.  
- Post-process outputs for clean question formatting.

## How to Get Started with the Model

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Zeeshan506/echohire-qgen-distilgpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Job Title: Backend Developer\nSkills: Python, FastAPI, PostgreSQL\nQuestions:"

inputs = tokenizer(prompt, return_tensors="pt")
input_ids = inputs["input_ids"]
attention_mask = (input_ids != tokenizer.pad_token_id).long()

output_ids = model.generate(
    input_ids=input_ids,
    attention_mask=attention_mask,
    max_new_tokens=300,
    num_beams=3,
    no_repeat_ngram_size=2
)

generated_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(generated_text)
```

## Training Details


### Training Data
- Dataset of ~100 jobs with 17–19 interview questions each.
- Input: Job title + skills
- Output: Structured questions

### Training Procedure
- Fine-tuned using Hugging Face Trainer API
- Epochs: 2
- Batch size: 2 per GPU
- Learning rate: 5e-5
- Tokenizer: distilgpt2 tokenizer, max input length 512
- Output: Structured questions

### Metrics

- **Training Loss:** Cross-entropy loss tracked during fine-tuning.  
- **Human Review:** Qualitative evaluation for relevance, coherence, and completeness of generated questions.  
- **Automated Evaluation (BERTScore):**  
  We performed BERTScore evaluation on a subset of the dataset (first 10 examples) to measure semantic similarity between the generated questions and reference completions using contextual embeddings.

  **BERTScore Results (first 10 examples):**  
  - **Precision:** 0.8691  
  - **Recall:** 0.8870  
  - **F1 Score:** 0.8780  

These results indicate that the model generates questions that are **highly aligned with reference completions**, confirming the quality and effectiveness of the fine-tuned model.

### Results

- Model successfully generates relevant interview questions for multiple technical domains.
- Some outputs may merge multiple questions; minor formatting post-processing recommended.

### Environmental Impact

- Hardware Type: NVIDIA Tesla T4 (Google Colab)
- Hours used: ~3–4 for training
- Compute Region: Colab USA/Europe
- Carbon Emitted: Minimal (single Colab GPU session)

### Technical Specifications

- Architecture: DistilGPT-2, causal language model
- Objective: Generate text conditioned on job title + skills
- Software: Transformers library, PyTorch, Hugging Face Hub

### Citation

If you use this model, please cite as:

#### APA:
Shah, S. Z. (2025). EchoHire Question Generator (DistilGPT-2 Fine-tuned). Hugging Face. 

(https://huggingface.co/Zeeshan506/echohire-qgen-distilgpt2)

#### BibTeX:
```
@misc{shah2025echohire,
  title={EchoHire Question Generator (DistilGPT-2 Fine-tuned)},
  author={Shah, Syed Zeeshan},
  year={2025},
  howpublished={\url{https://huggingface.co/Zeeshan506/echohire-qgen-distilgpt2}}
}
```