This model is a ModernBERT-base encoder with a two-layer projection head, fine-tuned with InfoNCE contrastive loss on the HALvest-Contrastive authorship attribution benchmark.

Interaction method: Patch-Level Late Interaction (PLI) with whole-word patches over ModernBERT embeddings. Subword tokens are grouped into whole-word patches using the tokeniser's word-boundary markers, then mean-compressed into a single vector per patch. MaxSim is computed between query and key patches. This is the best-performing PLI variant for cross-domain generalisation.

Intended use

Authorship attribution and authorship verification on English text. Given a query text and a set of candidate texts, the model scores how likely each candidate was written by the same author as the query.

How to use

This checkpoint is a PyTorch Lightning .ckpt file. You need the deep_stylometry codebase to load it.

Installation

bash

git clone https://github.com/Madjakul/deep_stylometry.git
cd deep_stylometry
pip install -r requirements.txt

Minimal inference

python

import torch
from transformers import AutoTokenizer
from deep_stylometry.modules.modeling_deep_stylometry import DeepStylometry
from deep_stylometry.utils.configs import BaseConfig

# 1. Load model from checkpoint
cfg = BaseConfig(mode="test").from_yaml("configs/test_pli_wholeword.yml")
model = DeepStylometry.load_from_checkpoint("last.ckpt", cfg=cfg)
model.eval()

# 2. Tokenize
tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-base")
texts = [
    "Query text whose authorship you want to identify.",
    "Candidate A: a text by the same author.",
    "Candidate B: a text by a different author.",
]
enc = tokenizer(
    texts, padding=True, truncation=True, max_length=512, return_tensors="pt",
)

# 3. Encode all texts through the model
with torch.no_grad():
    embs = model(enc["input_ids"], enc["attention_mask"])  # (3, seq_len, 768)

# 4. Score the query against each candidate
pool = model.contrastive_loss.pool  # the interaction module
with torch.no_grad():
    scores = pool(
        query_embs=embs[:1],
    key_embs=embs[1:],
    q_mask=enc['attention_mask'][:1],
    k_mask=enc['attention_mask'][1:],
    q_input_ids=enc['input_ids'][:1],  # word boundaries from token IDs
    k_input_ids=enc['input_ids'][1:],
    )
# scores shape: (1, 2) -- similarity of the query to [candidate A, candidate B]
# Higher score = more likely same author.
print(scores)

Training details

The model was trained on HALvest-Contrastive (English-only contrastive triplets derived from the HAL open-access scholarly repository) using InfoNCE loss with in-batch negatives. Training used 4x H100 GPUs with DDP, mixed precision (fp16), and a cosine learning rate schedule with linear warmup.

See the training configuration for full hyperparameters.

Citation

bibtex

@misc{kulumba_halvest_2026,
      title={HALvest-Contrastive: Retrieval-Like Authorship Attribution with Patch-Level Late Interaction},
      author={Francis Kulumba and Wissam Antoun and Guillaume Vimont and Laurent Romary and Florian Cafiero},
      year={2026},
      eprint={2407.20595},
      archivePrefix={arXiv},
      primaryClass={cs.DL},
      url={https://arxiv.org/abs/2407.20595},
}

bibtex

@misc{kulumba_does_2026,
      title={Where Does Authorship Signal Emerge in Encoder-Based Language Models?},
      author={Francis Kulumba and Guillaume Vimont and Laurent Romary and Florian Cafiero},
      year={2026},
      eprint={2605.19908},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.19908},
}

Madjakul/deep-stylometry-modernbert-pli-wholeword

deep-stylometry-modernbert-pli-wholeword

Model description