LessLM/person-name-match-likelihood-v6

Person Name Match Likelihood (v6)

Author: Elad Laor  ·  LinkedIn

A scoring head over MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli trained to predict whether two strings refer to the same person. Useful for record linkage, deduplication, and KYC-style identity matching where the only signal is a pair of name strings.

Quick start

from peft import PeftModel
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

BASE = "MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli"
ADAPTER = "LessLM/person-name-match-likelihood-v6"

tokenizer = AutoTokenizer.from_pretrained(BASE)
base = AutoModelForSequenceClassification.from_pretrained(BASE, num_labels=2)
model = PeftModel.from_pretrained(base, ADAPTER).eval()

def score(name_a: str, name_b: str) -> float:
    """Return P(same person) in [0, 1]."""
    inputs = tokenizer(name_a, name_b, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        logits = model(**inputs).logits
    return torch.softmax(logits, dim=-1)[0, 1].item()

print(score("John A. Smith", "J. Smith"))           # ~0.95 (initial expansion)
print(score("Yitzhak Cohen", "Itzhak Cohen"))       # ~0.99 (transliteration)
print(score("John Smith", "John Smyth"))            # ~0.90 (typo)
print(score("Robert Adams", "Roberta Adams"))       # ~0.05 (similar but different)

The model returns a 2-way softmax over [no_match, match]. The match probability is interpretable as a likelihood score; a temperature scaler (calibration.pt in this repo) is fitted on a held-out set if you want calibrated probabilities — load it and apply before the softmax for slightly tighter Expected Calibration Error.

Headline metrics

Evaluated on a held-out test set of 2,510 name pairs drawn from real public entity data (OpenSanctions) and a curated synthetic edge-case set.

Metric	Score
F1	0.9682
Precision	0.9568
Recall	0.9798
Accuracy	0.9733
Expected Calibration Error	0.0162
Latency (p95, CPU)	0.42 ms

Performance by edge case

Edge case	Accuracy	n
nickname (Bob ↔ Robert)	100.0%	121
name_order (Last, First ↔ First Last)	100.0%	112
transliteration (Yitzhak ↔ Itzhak)	100.0%	93
initial (J. Smith ↔ John Smith)	100.0%	112
middle_name add/drop	100.0%	50
title_suffix (Dr., Jr.)	100.0%	112
hyphenation, case_variation, combined, unrelated	100.0%	86
tricky_non_match (similar non-matches)	97.9%	331
partial_overlap	97.7%	353
unknown (real-world, no curated label)	97.5%	682
similar_name (Robert ↔ Roberta)	95.0%	341
typo	84.6%	117

The model is strongest on canonical edge cases (nicknames, initials, transliteration) and weakest on character-level typos where it overlaps with the similar_name distribution.

How it was trained

• Base: MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli (184M params, 12-layer encoder with disentangled attention).
• Adapter: LoRA (rank=16, alpha=32, dropout=0.1), targeting query_proj, key_proj, value_proj in every attention block. 600K trainable parameters (0.3% of base).
• Loss: Focal loss (γ=2.0) — down-weights easy examples, lets the model focus on hard pairs (similar names, typos).
• Optimizer: AdamW, LR=2e-4, cosine schedule, 10% warmup, weight decay 0.01.
• Schedule: 10 epochs, batch size 32, max sequence length 128 tokens, BF16 mixed precision.
• Seed: 42.
• Data: ~174K balanced match/no-match pairs (after 2.5× augmentation) — half drawn from OpenSanctions entities (real public records), half from a synthetic generator covering the 15 edge cases in the table above. Train/validation/calibration splits are entity-level (no person appears in more than one split) using a deterministic MD5-based hash so the splits reproduce bit-for-bit.

Bias, Risks, and Limitations

• Latin script only. The model was trained on Latin-script names. It will not work well on Hebrew, Arabic, Chinese, Cyrillic, etc. scripts unless they are first transliterated.
• OpenSanctions skew. The real-world half of the training data is drawn from a public sanctions/PEP entity database. Names in that distribution skew toward political, business, and criminal figures, with heavy representation of Russian, Ukrainian, Iranian, Chinese, and Latin American transliterations and a long tail of titles and honorifics. The model may behave differently on, say, US consumer-database names than on this distribution.
• Pair-level only. This is a pairwise matcher: given two name strings, score their likelihood of being the same person. It does not do blocking, clustering, or one-to-many matching. For dedup over a large list, pair it with a blocking layer (cheap pre-filter on first-letter, soundex, etc.) before invoking the model.
• Names alone. No surrounding context (DOB, email, address). Two real-world people with the same name will score as a match. Use this as one signal among several in a real identity-matching pipeline, not as the sole decision.
• Typo accuracy is the weakest cell. 84.6% on character-level typos. If your input is OCR output or hand-transcribed names, expect more errors in this category and consider a separate spell-correction step before scoring.
• No production guarantees. This is a research/portfolio artifact. Performance on your distribution may differ. Evaluate on a sample of your own data before relying on it.

Intended use

• Record linkage and dedup of person-name fields in datasets where you have only name strings to work with.
• KYC and identity-matching workflows as one feature among several.
• Benchmarking and research on encoder-based entity matching.

Out-of-scope use

• Non-Latin scripts (Hebrew, Arabic, Chinese, etc.) without prior transliteration.
• Surveillance, social scoring, or any use that would single out individuals for adverse treatment based on a name-match score alone.
• High-stakes one-shot identity decisions (eligibility, denial, arrest, eviction) — the model gives a likelihood, not a verdict.

License

MIT. You are free to use, modify, and redistribute, including commercially, provided you keep the attribution and license notice.

Citation

If you use this model, a backlink to this repo or the author's profile is appreciated.

@misc{laor2026_person_name_match_v6,
  author = {Elad Laor},
  title  = {Person Name Match Likelihood (v6) — a LoRA adapter on DeBERTa-v3 for pairwise person-name matching},
  year   = {2026},
  url    = {https://huggingface.co/LessLM/person-name-match-likelihood-v6}
}

Reproducibility & technical details

• Framework versions: peft==0.18.1, transformers>=4.40, torch>=2.0.
• Training environment: RunPod RTX 4090, ~3h wall-clock, BF16. Original v6 trained on RTX 3090 Ti (cross-GPU F1 delta: −0.0025).
• Seed: All randomness controlled by seed=42 (numpy, torch, transformers, dataloader generators). Re-running the training script with this seed and dataset version produces F1 within ±0.005 across BF16-capable GPUs.
• Calibration: calibration.pt is a single-parameter temperature scaler (T=0.95) fitted on a 1.5K held-out set. Apply it to logits before the final softmax to slightly reduce Expected Calibration Error from 0.016 to ~0.012.