NewsReX: A More Efficient Approach to News Recommendation with Keras 3 and JAX
Paper β’ 2508.21572 β’ Published
NRMS news recommendation model trained on MIND-small using the NewsReX framework (JAX).
| Seed | AUC | MRR | NDCG@5 | NDCG@10 |
|---|---|---|---|---|
| 42 * | 0.6614 | 0.3135 | 0.3469 | 0.4096 |
* Best seed (weights at repo root)
model:
name: nrms
architecture:
news_encoder:
type: multi_head_self_attention
num_heads: 12
head_dim: 15
attention_hidden_dim: 200
user_encoder:
type: multi_head_self_attention
num_heads: 20
head_dim: 15
attention_hidden_dim: 200
click_predictor:
type: dot_product
text_pooler:
type: attention
num_heads: 6
head_dim: 128
attention_query_dim: 200
dropout_rate: 0.0
embedding:
size: 300
trainable: true
dropout_rate: 0.2
seed: 42
inputs:
title:
max_length: 32
history:
max_length: 50
impressions:
max_length: 5
process_title: true
process_abstract: false
process_category: false
process_subcategory: false
process_user_id: false
training:
loss:
name: categorical_crossentropy
from_logits: true
reduction: sum_over_batch_size
label_smoothing: 0.0
optimizer: adam
learning_rate: 0.0001
batch_size: 256
num_epochs: 10
early_stopping:
patience: 3
min_improvement: 0.01
negative_sampling:
strategy: random
candidates: 4
evaluation:
mode: fast
metrics:
- auc
- mrr
- ndcg@5
- ndcg@10
batch_size: 256
newsrex/NRMS-JAX-MIND-small-bert-random/
βββ model.safetensors β best seed (42)
βββ test_results.json
βββ training_run_summary.json
βββ seed_42/model.safetensors
βββ README.md
git clone https://github.com/igor17400/NewsReX.git
cd NewsReX && uv sync
# Run evaluation with best seed weights
uv run python src/eval.py \
experiment=mind/glove/nrms \
framework=jax \
weights=hf://newsrex/NRMS-JAX-MIND-small-bert-random/model.safetensors
# Run evaluation with a specific seed
uv run python src/eval.py \
experiment=mind/glove/nrms \
framework=jax \
weights=hf://newsrex/NRMS-JAX-MIND-small-bert-random/seed_42/model.safetensors
@misc{newsrex2026,
title={NewsReX: An Open-Source Multi-Framework for Neural News Recommendation},
author={Igor L. R. Azevedo and Toyotaro Suzumura and Yuichiro Yasui},
year={2025},
eprint={2508.21572},
archivePrefix={arXiv},
primaryClass={cs.IR},
url={https://arxiv.org/abs/2508.21572},
}