Update documentation

README.md

@@ -12,26 +12,15 @@ task_ids:
 
 # abstracts-embeddings
 
-This is the embeddings of the titles and abstracts of 95 million academic publications taken from the [OpenAlex](https://openalex.org) dataset as of May 5, 2023. The script that generated the embeddings is available on [Github](https://github.com/colonelwatch/abstracts-search/blob/master/build.py), but the general process is as follows:
+This is the embeddings of the titles and abstracts of 110 million academic publications taken from the [OpenAlex](https://openalex.org) dataset as of January 1, 2025. The embeddings are generated with a Unix pipeline, chaining together the AWS CLI, gzip, `oa_jsonl` (a C parser tailored to the JSON Lines structure of the OpenAlex snapshot), and a Python embedding script. The source code of `oa_jsonl` and the Makefile which sets up the pipeline is available on [Github](https://github.com/colonelwatch/abstracts-search), but the general process is as follows:
 
-1. Reconstruct the text of the abstract from the inverted index format
-2. Construct a single document string in the format `title + ' ' + abstract` or just `abstract` if there is no title
-3. Determine if the document string is in English using [fastText](https://fasttext.cc/docs/en/language-identification.html)
-4. If it is in English, compute an embedding using the `all-MiniLM-L6-v2` model provided by [sentence-transformers](https://www.sbert.net/)
+1. Decode the JSON entry of an individual work
+2. From the language field, determine if the abstract will be in English, and if not, go back to step 1
+3. From the abstract inverted index field, reconstruct the text of the abstract
+4. If there is a title field, construct a single document in the format `title + ' ' + abstract`, or if not, just use the abstract
+5. Compute an embedding with the [stella_en_1.5B_v5](https://huggingface.co/NovaSearch/stella_en_1.5B_v5) model
+6. Write it to a local SQLite3 database
 
-Though the OpenAlex dataset records 240 million works, not all of these works have abstracts or are in English. However, the `all-MiniLM-L6-v2` model was only trained on English texts, hence the filtering.
+Said database is then exported in parquet format as pairs of OpenAlex IDs and length-1024 float32 vectors. The model was run with bfloat16 quantization, yielding bfloat16 vectors, but the conversion from bfloat16 to float32 leaves the lower two bytes as all-zero. This was exploited with byte-stream compression to store the vectors in a parquet with full precision but no wasted space. This does however mean that opening the parquets in the Hugging Face `datasets` library will lead to the cache using twice the space.
 
-## Dataset Structure
-
-In the future, this dataset might become a parquet in order to admit all the features offered by Hugging Face Datasets, but it consists only of a text file and a numpy memmap for now. The memmap is an array of many length-384 `np.float16` vectors, and the i-th row vector in this array corresponds with the i-th line in the text file. The text file is just a list of ids that can be used to get more information from the OpenAlex API.
-
-```python
-import numpy as np
-
-with open('openalex_ids.txt', 'r') as f:
-    idxs = f.read().splitlines()
-
-embeddings = np.memmap('embeddings.memmap', dtype=np.float16, mode='r').reshape(-1, 384)
-```
-
-However, the memmap cannot be uploaded to Hugging Face as a single file, so it's split with the command `split -b 3221225472 -d --suffix-length=3 --additional-suffix=.memmap embeddings.memmap embeddings_`. It can be put back together with the command `cat embeddings_*.memmap > embeddings.memmap`.
+Though the OpenAlex dataset records 240 million works, not all of these works have abstracts or are in English. Besides the works without abstracts, the stella_en_1.5B_v5 model was only trained on English texts, hence the filtering.