Open science
Open models, benchmarks and data
What we give back to the AI community: models, benchmarks, annotated data and code, each with its paper. Figures from Hugging Face are refreshed daily.
Benchmark and models
PatenTEB and patembed
A benchmark of 15 tasks and 2.06 million examples for patent text embedding, with the patembed family of models. patembed-base reports 0.494 on the MTEB BigPatentClustering.v2 task, against 0.445 for the previous best; the models are not yet submitted to the MTEB leaderboard.
Dataset, part of MTEB
DAPFAM
A family-level dataset to benchmark cross-domain patent retrieval. Its 18 tasks, and a new “patent retrieval” task type, were merged into MTEB, the reference benchmark for text embeddings, in September 2025.
Expert-annotated data
Contradictions in patents
1,600 patents containing a TRIZ contradiction and 1,600 without, selected by experts from about 20,000 read, to train and evaluate contradiction mining (IEEE ICDM 2021).
Hugging Face
Models
Patent text embedding models, trained in Strasbourg on a frugal budget.
Models and datasets are published under the account of Iliass Ayaou (datalyes); the Chair’s organisation on Hugging Face is INSAstrasbourg. huggingface.co/datalyes · huggingface.co/INSAstrasbourg
12 models · 23,156 downloads in total · licence CC BY-NC-SA 4.0 (non-commercial)
| Name | Task | Downloads | Likes |
|---|---|---|---|
| patembed-large | sentence-similarity | 13,901 | 2 |
| patembed-base | sentence-similarity | 6,176 | 0 |
| patembed-base_long_4096 | sentence-similarity | 2,025 | 0 |
| patembed-base_small | sentence-similarity | 578 | 0 |
| patembed-large_no_prompts | sentence-similarity | 476 | 0 |
| patembed-small | sentence-similarity | 0 | 0 |
| patembed-mini | sentence-similarity | 0 | 0 |
| patembed-nano | sentence-similarity | 0 | 0 |
| patembed-base_long_1024 | sentence-similarity | 0 | 0 |
| patembed-base_long_2048 | sentence-similarity | 0 | 0 |
| patembed-large_all_ret_only | sentence-similarity | 0 | 0 |
| patembed-large_all_no_classif | sentence-similarity | 0 | 0 |
Hugging Face
Datasets and benchmarks
Retrieval, classification and clustering tasks built from patents, ready to reuse.
16 datasets · 11,720 downloads in total
| Name | Downloads | Likes |
|---|---|---|
| DAPFAM_patent | 8,307 | 9 |
| retrieval_IN | 348 | 1 |
| problem2full | 304 | 0 |
| effect2full | 295 | 0 |
| title2full | 249 | 0 |
| problem2solution | 242 | 1 |
| retrieval_OUT | 230 | 1 |
| clusters_inventor | 225 | 0 |
| effect2substance | 218 | 0 |
| retrieval_MIXED | 204 | 1 |
| clusters_ext_full_ipc | 203 | 0 |
| class_bloom | 190 | 0 |
| para_solution | 188 | 0 |
| class_text2ipc3 | 181 | 0 |
| class_nli_oldnew | 173 | 0 |
| para_problem | 163 | 0 |
Code
Public repositories accompanying the papers. The flagship tools themselves are reserved for members.
- PatenTEB ↗
Code and resources of the PatenTEB benchmark.
- PaGAN ↗
Contradiction mining in patents (IEEE ICDM 2021), with the annotated dataset.
- MTEB — DAPFAM tasks ↗
Our contribution to the reference embedding benchmark.
GPU hours, cloud credits, annotated problems or expert time to share?
Contribute resources