| --- |
| license: mit |
| task_categories: |
| - feature-extraction |
| - text-retrieval |
| tags: |
| - embeddings |
| - vector-database |
| - retrieval |
| - beir |
| - la2m |
| pretty_name: LA2M Embeddings |
| --- |
| |
| # LA2M_emb |
| |
| Embedding vectors used in the paper |
| **[Integrating Vector Databases across Embedding Models](https://dl.acm.org/doi/10.1145/3769803)** |
| (Beining Yang, Yang Cao, Yang Ren. *Proc. ACM Manag. Data* 3(6) (SIGMOD), Article 338, 2025). |
| |
| This is a **curated subset of [DB-Edinburgh/VectorBenchmark](https://huggingface.co/datasets/DB-Edinburgh/VectorBenchmark)**, |
| restricted to exactly the 6 embedding models (Table 3) and 5 benchmarks (Table 2) reported in the paper, |
| so that the experiments can be reproduced without pulling the full ~6k-file benchmark. |
| |
| Code: [VectorMerge](https://github.com/Suchun-sv/VectorMerge) (`reproduce` branch has a Docker setup). |
| |
| ## Contents |
| |
| `embeddings/{corpus,query}_embeddings_{model}_{dataset}.npy` — plain `numpy` arrays, one row per document / query, |
| row order matches the BEIR `corpus.jsonl` / `queries.jsonl` of the corresponding dataset. |
|
|
| | paper model (Table 3) | file key | dim in these files | dim in Table 3 | |
| |---|---|---|---| |
| | GloVe | `glove` | 300 | 300 | |
| | FastText | `fast-text` | 300 | 300 | |
| | Mistral | `mistral` | 1024 | 1024 | |
| | NV-Embed-V2 | `nv-embed` | 4096 | 1024 | |
| | GTE-Qwen2 | `gte` | 3584 | 1024 | |
| | OpenAI-Ada | `openai` | 1536 | 1536 | |
|
|
| | dataset | corpus rows | query rows | |
| |---|---|---| |
| | `scifact` | 5,183 | 300 | |
| | `nfcorpus` | 3,633 | 323 | |
| | `arguana` | 8,674 | 1,406 | |
| | `scidocs` | 25,657 | 1,000 | |
| | `fiqa` | 57,638 | 648 | |
|
|
| Corpus row counts match Table 2 of the paper exactly. |
|
|
| ## Known quirks |
|
|
| These are inherited from the source benchmark and are documented rather than silently fixed: |
|
|
| - **`nv-embed` vs `nv-embed-v2`.** The paper's NV-Embed-V2 vectors are stored under the `nv-embed` key, |
| which is also the name the `vectormerge` code expects. `nv-embed-v2`-keyed files exist only for |
| `scidocs` and `fiqa`; they are included here for completeness and have identical shapes. |
| - **`query_embeddings_mistral_scifact.npy` has 320 rows**, not the 300 that every other model has for |
| SciFact (and that Table 2 reports). Index-aligned evaluation on that one file needs care. |
| - **Mixed dtypes.** Most files are `float32`; `openai`, `mistral`, and `glove` (arguana, fiqa) are `float64`. |
| - **Reported vs stored dimension.** Table 3 lists 1024 for NV-Embed-V2 and GTE-Qwen2, while the stored |
| vectors are at native dimension (4096 / 3584). The `vectormerge` pipeline aligns dimensions at load time |
| (`align_dimension`), so the 1024 in the table refers to the model family rather than these files. |
| |
| ## Usage |
| |
| ```python |
| from huggingface_hub import hf_hub_download |
| import numpy as np |
| |
| path = hf_hub_download( |
| "DB-Edinburgh/LA2M_emb", |
| "embeddings/corpus_embeddings_mistral_scifact.npy", |
| repo_type="dataset", |
| ) |
| X = np.load(path) # (5183, 1024) |
| ``` |
| |
| ## Citation |
| |
| ```bibtex |
| @article{Yang2025integrating, |
| author = {Beining Yang and Yang Cao and Yang Ren}, |
| title = {Integrating Vector Databases across Embedding Models}, |
| journal = {Proc. {ACM} Manag. Data}, |
| volume = {3}, |
| number = {6}, |
| pages = {1--28}, |
| year = {2025}, |
| doi = {10.1145/3769803} |
| } |
| ``` |
| |
