Resumator v2

Fine-tuned sentence-transformer for resume-to-job matching. Encodes resumes and job descriptions into embeddings for cosine similarity search.

Trained with Matryoshka Representation Learning — supports truncation to 384/256/128 dims at inference with minimal quality loss.

What's New in v2

Usage

from sentence_transformers import SentenceTransformer

# Load with Matryoshka truncation to 384 dims
model = SentenceTransformer("shankerram3/resumator", truncate_dim=384)

candidate = "Name: Jane Doe\nSkills: Python, React, PostgreSQL\nResume: Full-stack engineer with 4 years..."
job = "Title: Senior Software Engineer\nCompany: TechCorp\nDescription: Looking for full-stack engineer..."

embeddings = model.encode([candidate, job], normalize_embeddings=True)
similarity = float(embeddings[0] @ embeddings[1])  # cosine similarity

Input Format

Candidate:

Name: {name}
Titles: {title1}, {title2}
Skills: {skill1}, {skill2}, {skill3}
Experience: {years} years
Location: {city}, {state}
Resume: {full_resume_text}

Job:

Title: {job_title}
Company: {company_name}
Location: {city}, {state}
Experience: {years} years
Description: {full_job_description}

Benchmarks

Held-Out Evaluation (Unseen Data)

Evaluated on 500 resume-job pairs using jobs the model never saw during training, scored by Cerebras gpt-oss-120b as ground truth. Run on AMD Instinct MI300X GPU.

• Spearman: Rank correlation with LLM recruiter scores (higher = better ranking)
• Ranking Acc: % of (good, bad) pairs where the good match scores higher
• Separation: Mean cosine sim gap between good matches (≥0.6) and bad matches (≤0.3)
• IQR: Interquartile range — higher means better discrimination between matches

MTEB Benchmarks

Standard sentence-transformer evaluation on public benchmarks (not domain-specific):

v2 improves on 11 of 13 MTEB tasks despite being fine-tuned for a specific domain.

Speed

MI300X GPU:

Apple M-series CPU:

Score Distribution (2,500 resume-job pairs)

v2 distribution:
0.0-0.1:                                             13 (  0.5%)
0.1-0.2: ########                                   149 (  6.0%)
0.2-0.3: ################################           590 ( 23.6%)
0.3-0.4: ########################################   722 ( 28.9%)
0.4-0.5: ##############################             543 ( 21.7%)
0.5-0.6: ######################                     410 ( 16.4%)
0.6-0.7: ###                                         72 (  2.9%)
0.7-0.8:                                              1 (  0.0%)

v2 spreads scores more widely (IQR 0.186 vs 0.133), making it easier to set thresholds and distinguish good matches from mediocre ones.

Matryoshka Dimension Quality

Trained with Matryoshka Representation Learning — quality at different truncation levels:

384-dim truncation actually improves separation — Matryoshka loss concentrates the most discriminative information in the first 384 dimensions.

Training Details

Data

• 15,600 resume-job pairs scored 0.0–1.0 by Cerebras gpt-oss-120b
• 51 candidates × 300 random jobs each
• Scoring criteria: technical skills overlap (50%), role/title alignment (30%), experience level match (20%)

Training Pipeline

1. Phase 1: MatryoshkaLoss wrapping MultipleNegativesRankingLoss (5 epochs)
   • Triplets: anchor (candidate) + positive (score ≥ 0.6) + hard negative (score ≤ 0.3)
   • 2,702 triplets, effective batch size 32 (8 × 4 gradient accumulation)
   • Matryoshka dims: [768, 384, 256, 128]
2. Phase 2: CosineSimilarityLoss refinement (2 epochs)
   • Mid-range pairs (0.3 < score < 0.6) for calibrating the middle of the score range
   • 2,503 pairs, learning rate 5e-6

Training Logs

Phase 1 — MatryoshkaLoss + MNRL (5 epochs):

Phase 2 — CosineSimilarityLoss (2 epochs):

Hardware

• Training: AMD Instinct MI300X (ROCm 6.2), ~3 minutes total
• Evaluation: AMD Instinct MI300X + Apple M-series CPU

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 5.2.3
• Transformers: 5.3.0
• PyTorch: 2.5.1+rocm6.2
• Accelerate: 1.13.0
• Datasets: 4.7.0

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 384, 'do_lower_case': False, 'architecture': 'MPNetModel'})
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_mean_tokens': True})
  (2): Normalize()
)

Why Fine-Tune?

Generic sentence-transformers treat resumes as arbitrary text. Fine-tuning teaches:

• Domain vocabulary: "OPT", "H1B", "C2C" are visa types, not random acronyms
• Structural alignment: Match skills sections to requirements sections
• Experience calibration: "3 years Java" → "mid-level Java developer", not "senior architect"
• Description reading: v2 matches on actual job description content, not just title keywords

General-purpose models (bge, e5) score 0.19–0.24 Spearman on resume-job matching despite leading MTEB leaderboards. Domain fine-tuning is essential for this task.

Limitations

• English only, US tech market bias
• 512 token limit — key information should appear early in the text
• Trained on tech/IT recruiting data — may underperform on non-tech roles
• LLM-scored training data inherits any biases from the scoring model

Citation

BibTeX

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

License

Apache 2.0