dhruvi5319/llama-3.2-3b-resume-fit-summary

Llama 3.2 3B — Resume Fit-Summary (LoRA)

A QLoRA adapter for meta-llama/Llama-3.2-3B-Instruct that generates a structured 4-point resume↔job-description fit summary: (1) experience summary, (2) key strengths, (3) weaknesses/missing qualifications, (4) overall fit verdict.

Built to replace the GPT-3.5 API call in Smart Resume Screener with a free, local, consistent model. Trained by distillation from a GPT-4-class teacher on the screener's exact prompt.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "meta-llama/Llama-3.2-3B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained(base, device_map="auto"),
    "dhruvi5319/llama-3.2-3b-resume-fit-summary",
).eval()

prompt = """You are an AI assistant tasked with evaluating a candidate's resume against a job description.
1. Summarize the candidate's experience and qualifications.
2. Identify their key strengths from the resume.
3. Identify possible weaknesses or missing qualifications based on the job description.
4. Evaluate their overall fit for the position.

Resume:
{resume}

Job Description:
{job}

Provide a concise paragraph covering the above 4 points."""

msgs = [{"role": "user", "content": prompt.format(resume=RESUME, job=JD)}]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True,
                             return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**enc, max_new_tokens=300, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

Training

• —Method: QLoRA (4-bit NF4 base + LoRA r=16, α=32) with completion-only loss.
• —Data: 471 distilled (resume+JD → ideal summary) examples — real pairs from cnamuangtoun/resume-job-description-fit plus synthetic strong-fit augmentation; teacher = gpt-4o-mini.
• —Schedule: 3 epochs, LR 2e-4 cosine, effective batch 16, on a single A100.

Evaluation (held-out, n=30, GPT-4-class judge, position-randomized)

Limitations & intended use

• —Scoped to English resume/JD fit summaries in the 4-point format above.
• —The base model is already a capable summarizer, so gains are in consistency, conciseness, and format-adherence, not raw capability.
• —May reflect biases in the underlying resume dataset; not a substitute for human review in hiring.