domynom/resume-job-deberta

Resume-Job Fit Classifier (DeBERTa-v3)

microsoft/deberta-v3-small fine-tuned as a 3-class sequence-pair classifier that scores how well a resume fits a job description. Built for Northeastern IE 7500 (Applied NLP). Powers the Resume Job Fit Matcher Space.

Usage

Pass the resume as the first sequence and the job description as the second.

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tok = AutoTokenizer.from_pretrained("domynom/resume-job-deberta")
model = AutoModelForSequenceClassification.from_pretrained("domynom/resume-job-deberta")

inputs = tok(resume_text, job_text, truncation=True, max_length=512,
             return_tensors="pt")
pred = model(**inputs).logits.argmax(-1).item()
print(model.config.id2label[pred])

Labels

Label order is authoritative in config.json — read id2label rather than assuming positional order:

• 0 -> No Fit
• 1 -> Potential Fit
• 2 -> Good Fit

Training

Fine-tuned on cnamuangtoun/resume-job-description-fit. The dataset's own test split is held out for final evaluation; a 90/10 split of train supplies validation.

Setting	Value
Base model	microsoft/deberta-v3-small
Learning rate	1e-5
Epochs	8
Batch size	4
Gradient accumulation	2
Precision	bf16
Weight decay	0.01
Max sequence length	512
Seed	52

Reproduce with python scripts/train_deberta.py from the project repo.

Evaluation

Held-out test split of the same dataset:

Metric	Score
Accuracy	0.4974
Macro F1	0.3946
Weighted F1	0.4561
Macro precision	0.4191
Macro recall	0.4077

Limitations

This checkpoint is published as the architecture-comparison point in the project's evaluation, not as its recommended scorer. It scores below the DistilBERT sibling (0.395 vs 0.432 macro-F1) and leans hard on the majority class, recalling only 15.5% of Potential Fit. An ordinal (CORAL) variant of the same base reaches 0.444 macro-F1 with much better balance; see the project's ablation study.

The DeBERTa-v3 tokenizer needs sentencepiece installed.

Fit on this dataset is genuinely hard and the scores above are close to the three-class floor — treat a single prediction as a weak signal, not a verdict. In the project's own evaluation a TF-IDF baseline matched or beat every transformer here on macro-F1. The model inherits whatever role, seniority, and demographic bias exists in the underlying postings, and it has not been audited for disparate impact. Do not use it to screen real candidates.