README.md

---
base_model: unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit
library_name: peft
license: apache-2.0
pipeline_tag: text-generation
tags:
- lora
- qlora
- sft
- transformers
- trl
- unsloth
- resume-screening
- recruitment
- information-extraction
- json-generation
- hiring
---

# Model Card for HireSense Resume Parser LoRA

## Model Details

### Model Description

HireSense Resume Parser LoRA is a fine-tuned adapter model built on top of Qwen3-4B-Instruct using QLoRA and supervised fine-tuning (SFT). The model is designed to extract structured JSON information from resumes for downstream recruitment and candidate-job matching workflows.

The model converts raw resume text into a consistent structured schema containing:
- Personal information
- Skills
- Education
- Work experience
- Projects
- Certifications

This model is intended to be used as a component in AI-powered hiring pipelines and resume analysis systems.

- **Developed by:** Rohit BK
- **Model type:** Causal Language Model (LoRA Adapter)
- **Language(s):** English
- **License:** Apache-2.0
- **Finetuned from model:** unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit

---

## Model Sources

- **Base Model:** Qwen3-4B-Instruct
- **Frameworks:** Transformers, PEFT, TRL, Unsloth

---

# Uses

## Direct Use

This model is intended for:
- Resume parsing
- Structured information extraction
- Candidate profile generation
- Resume-to-JSON conversion
- Recruitment automation systems

Example output schema:

```json
{
  "name": "John Doe",
  "email": "john@example.com",
  "phone": "9876543210",
  "skills": ["Python", "React", "SQL"],
  "education": [
    {
      "degree": "B.Tech",
      "institution": "XYZ University",
      "year": "2025"
    }
  ]
}
```

---

## Downstream Use

The model can be integrated into:
- Applicant Tracking Systems (ATS)
- Resume ranking systems
- Semantic candidate matching pipelines
- Recruitment copilots
- Hiring analytics dashboards

---

## Out-of-Scope Use

This model is NOT intended for:
- Final hiring decisions
- Automated candidate rejection without human review
- Personality assessment
- Predicting candidate success
- Sensitive demographic inference

Human oversight is strongly recommended.

---

# Bias, Risks, and Limitations

The model may:
- Produce incorrect or incomplete JSON
- Miss information in poorly formatted resumes
- Exhibit biases inherited from training data
- Struggle with multilingual resumes
- Perform inconsistently on highly creative resume layouts

The model should not be used as the sole decision-maker in hiring processes.

---

# Recommendations

Users should:
- Validate generated outputs before use
- Use human review for hiring decisions
- Combine the model with rule-based validation systems
- Avoid relying solely on generated scores or rankings

---

# How to Get Started with the Model

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen3-4B-Instruct"
adapter_id = "YOUR_USERNAME/HireSense-ResumeParser-LoRA"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = """
Extract structured JSON information from the following resume.

Resume:
John Doe
Python Developer
Skills: Python, React, SQL
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256
)

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

---

# Training Details

## Training Data

The model was trained on structured resume-to-JSON instruction pairs containing:
- Resume text
- Extraction prompts
- Structured JSON outputs

Training data included synthetic and manually curated resume samples.

---

## Training Procedure

The model was fine-tuned using:
- QLoRA
- Supervised Fine-Tuning (SFT)
- 4-bit quantization
- PEFT adapters

### Training Hyperparameters

- **Training regime:** bf16 mixed precision
- **Fine-tuning method:** QLoRA
- **Quantization:** 4-bit NF4
- **Optimizer:** AdamW
- **Frameworks:** Transformers + TRL + Unsloth

---

# Evaluation

## Metrics

The model was evaluated qualitatively on:
- JSON validity
- Field extraction accuracy
- Structural consistency
- Hallucination frequency

---

## Results

The model demonstrated:
- Consistent JSON generation
- Good extraction performance on structured resumes
- Improved formatting consistency compared to the base model

Performance may degrade on:
- Image-based resumes
- Multi-column layouts
- Highly unstructured resumes

---

# Environmental Impact

- **Hardware Type:** NVIDIA GPU
- **Training Framework:** Unsloth
- **Quantization:** 4-bit QLoRA

---

# Technical Specifications

## Model Architecture and Objective

This model uses:
- Qwen3-4B-Instruct as the base model
- LoRA adapters for parameter-efficient fine-tuning
- Causal language modeling objective

---

# Citation

## BibTeX

```bibtex
@misc{hiresense2026,
  title={HireSense Resume Parser LoRA},
  author={Rohit BK},
  year={2026},
  publisher={Hugging Face}
}
```

---

# Model Card Authors

Rohit BK

---

# Model Card Contact

For questions or collaboration inquiries, please contact through Hugging Face or GitHub.

---

### Framework versions

- PEFT 0.19.1
- Transformers
- TRL
- Unsloth