Math journal suggester

About

An experiment in recommending math journals from a paper’s title and abstract using a fine-tuned Kev-4B model.

How recommendations work

Kev scores all 95 journal names from the manuscript’s title and abstract. The five recommendations are numbered 1–5 in descending score order. Journal names appear in a fixed alphabetical order inside the model’s request.

Kev-4B is hosted on an ASUS Ascent GX10, based on NVIDIA DGX Spark.

Data collection

The dataset contains papers published from 2016 through 2025 in 95 journals and series. Training uses 8,000 papers, with separate sets of 1,000 validation and 1,000 test papers. Sampling approximately follows publication counts. Validation and test each include at least one paper per journal.

Publication records and available abstracts come from Crossref, OpenAlex, publishers, and arXiv, with identity checks before matching.

IHÉS, Annals, Acta Mathematica, Inventiones, and JAMS account for 105 training papers and 18 validation papers. Limited training coverage may contribute to weaker performance.

Test results

We compared released Kev, before our journal fine-tuning, with the fixed epoch-2.5 checkpoint. Both scored the same 95 journals for the same 1,000 test papers on a B300. Top 1, Top 3, and Top 5 measure whether the actual publication journal appears among the first one, three, or five recommendations.

ModelTop 1Top 3Top 5
Released Kev15.7%29.8%37.3%
Fine-tuned Kev57.2%78.9%85.3%

Fine-tuning improved Top 3 accuracy by 49.1 percentage points.

Test papers were excluded from training and checkpoint selection. The benchmark had been examined during earlier project development.

Learning-rate comparison

Each trial continued training released Kev’s LoRA adapter and decision head on the same 1,000 papers for two epochs. All trials used seed 42 and the same 1,000 validation papers.

Peak rateEpoch 1 Top 3Epoch 2 Top 3
1e-548.0%49.7%
2e-549.8%51.0%
4e-551.9%51.4%

We selected 4e-5 by validation Top 3. This comparison uses one seed and does not establish an optimal rate.

8,000-paper training results

The full run started again from released Kev at a peak learning rate of 4e-5, using AdamW and a five-epoch OneCycle schedule. Each optimizer update accumulated eight single-paper batches. Validation ran every half epoch.

After at least two epochs, three checks without a new best validation Top 3 score triggered early stopping. Training stopped at epoch 4 and selected epoch 2.5.

EpochValidation Top 3Loss
0.556.3%2.560
170.0%1.942
1.572.7%1.826
275.4%1.686
2.5 (selected)76.2%2.045
376.2%1.837
3.574.5%2.700
474.7%3.398

Loss is cross-entropy over all 1,000 validation papers. Selection used Top 3 accuracy, so the chosen checkpoint does not have the lowest loss.