resume-screener-lora

resume-screener-lora

A LoRA adapter for Qwen2.5-0.5B-Instruct that returns structured JSON verdicts for resume screening by default, without prompt engineering.

Code, training notebook, evaluation and hosted demo: github.com/ayush-s-tomar/resume-screener-lora

Output

For a prompt such as Screen this resume for a Backend Engineer position and return a structured verdict. followed by the resume, the model returns:

{
  "role": "Backend Engineer",
  "ats_score": 56,
  "verdict": "moderate_match",
  "matched_skills": ["Python", "PostgreSQL", "Docker", "Kubernetes"],
  "missing_skills": ["Java", "REST APIs", "Redis"],
  "years_experience": 4
}

Verdicts follow the training data: strong_match (ats_score 75-100), moderate_match (50-74) and weak_match (8-48).

Usage

import json
import re
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen2.5-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, "Kus-hal/resume-screener-lora")

prompt = (
    "Screen this resume for a Backend Engineer position and return a structured verdict.\n\n"
    "Resume: 4 years of backend experience. Skilled in Python, PostgreSQL, Docker, Kubernetes."
)
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True,
)
output = model.generate(**inputs, max_new_tokens=200, do_sample=False)
text = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(json.loads(re.search(r"\{.*\}", text, re.DOTALL).group(0)))

Training

• Method: LoRA, r=16, alpha=32, dropout 0.05, on the q/k/v/o projections (2.16M of 496M parameters, 0.44%)
• Data: 800 train / 96 eval examples, synthetic, generated by data/generate_dataset.py in the GitHub repo
• Hardware: 3 epochs on a free Colab T4

Evaluation (96 held-out examples)

Limitations

• The labels come from the repository's synthetic dataset generator, so accuracy measures agreement with that rubric, not with human recruiters.
• The eval set is small (96 examples), so treat the numbers as indicative.
• On six hand-written resumes the adapter ranked them sensibly and rejected a clear mismatch, but it was more conservative than a larger prompted model (it never returned strong_match there).
• This is a portfolio and research project, not a tool for making hiring decisions.