| license: mit | |
| # PoC (Patents with One Citation) dataset | |
| This dataset is useful for training or evaluating models that predict patent-to-patent similarity, such as those used for patent searching. | |
| It was developed and used for the training of an ML model that powers the [PQAI](https://search.projectpq.ai/) search engine. | |
| ## Details | |
| The dataset contains 90,013 samples. | |
| Each sample contains: | |
| - a subject patent (`sp`) | |
| - its only citation (`cit`) | |
| - its CPC code (`cpc`) | |
| - a list of 10 patents (`sims`) that are similar to `sp` (in that they share the CPC code) and published before `sp` | |
| Every line of the dataset is a JSON parsable string (`.jsonl` format), which upon parsing given an array of this format: | |
| ``` | |
| [pn, cit, cpc, [...sims]] | |
| ``` | |
| ## Task | |
| Given the subject patent `sp` the task is to assign a similarity score to each patent `[cit, ...sims]`. Ideally, the score should be maximum for `cit`. | |
| ## Metrics | |
| It's a ranking task, so the following metrics make the most sense: | |
| - DCG/NDCG | |
| - Accuracy |