Read me is ready

README.md

@@ -0,0 +1,44 @@
+Output:
+
+``{<br/>'cat': ['mammal', 'animal'], <br/> 'dog': ['hound', 'animal'], <br/>'economics and sociology': ['both fields of study'], <br/>'public company': ['company']<br/>}
+
+### How was it trained?``
+
+1. Using Google's T5-base and T5-small. Both models are released on the Hugging Face Hub.
+2. T5-base was trained for only two epochs while T5-small was trained for 5 epochs.
+
+## Where did you get the data?
+
+1. I extracted and curated a fragment of [Conceptnet](https://conceptnet.io/)
+2. In particular, only the IsA relation was used.
+3. Note that one thing can belong to multiple concepts (which is pretty cool if you think about [Fuzzy Description Logics](https://lat.inf.tu-dresden.de/~stefborg/Talks/QuantLAWorkshop2013.pdf)).
+Multiple inheritances however mean some terms belong to so many concepts. Hence, I decided to randomly throw away some due to the **maximum length limitation**.
+
+### Setup
+1. I finally allowed only `2` to `4` concepts at random for each term. This means, there is still great potential to make the models generalise better 🚀.
+3. I used a total of `279884` training examples and `1260` for testing. Edges -- i.e `IsA(concept u, concept v)` -- in both sets are disjoint.
+4. Trained for `15K` steps with learning rate linear decay during each step. Starting at `0.001`
+5. Used `RAdam Optimiser` with weight_decay =`0.01` and batch_size =`36`.
+6. Source and target max length were both `64`.
+
+### Multilingual Models
+
+1. The "conceptor" model is multilingual. English, German and French is supported.
+2. [Conceptnet](https://conceptnet.io/) supports many languages, but I just chose those three because those are the ones I speak.
+
+### Metrics for flexudy-conceptor-t5-base
+
+| Metric        |         Score |
+| ------------- |:-------------:|
+| Exact Match   | 36.67         |
+| F1            | 43.08         |
+| Loss smooth   | 1.214         |
+
+Unfortunately, we no longer have the metrics for flexudy-conceptor-t5-small. If I recall correctly, base was just slightly better on the test set (ca. `2%` F1).
+
+## Why not just use the data if you have it structured already?
+
+Conceptnet is very large. Even if you just consider loading a fragment into your RAM, say with only 100K edges, this is still a large graph.
+Especially, if you think about how you will save the node embeddings efficiently for querying.
+If you prefer this approach, [Milvus](https://github.com/milvus-io/pymilvus) can be of great help.
+You can compute query embeddings and try to find the best match. From there (after matching), you can navigate through the graph at `100%` precision.