Datasets:
_id stringlengths 21 254 | embedding list |
|---|---|
angular-cli/CODE_OF_CONDUCT.md_0_3649 | [
-0.0336763896048069,
-0.022734761238098145,
-0.006464448757469654,
-0.026947373524308205,
0.008758424781262875,
0.012756149284541607,
0.030198780819773674,
-0.027848459780216217,
-0.014526035636663437,
-0.011967242695391178,
-0.01671210676431656,
0.01681799814105034,
0.01734916865825653,
-... |
angular-cli/LICENSE_0_1105 | [
-0.02577986754477024,
-0.04667440056800842,
-0.03516227379441261,
0.01505901850759983,
0.003491516225039959,
0.00765308178961277,
0.0013274888042360544,
0.028498204424977303,
0.020353222265839577,
-0.003634824650362134,
0.028108498081564903,
0.06382276117801666,
0.03745859116315842,
-0.013... |
angular-cli/WORKSPACE_0_4925 | [
0.004350633826106787,
-0.01684643141925335,
-0.017954882234334946,
-0.014836742542684078,
0.013532682321965694,
-0.051696013659238815,
0.016014328226447105,
0.014475037343800068,
-0.011623299680650234,
0.0229511596262455,
0.025274287909269333,
0.011875072494149208,
-0.00854075700044632,
0.... |
angular-cli/WORKSPACE_4927_8300 | [
-0.01590995118021965,
-0.028426384553313255,
-0.026947246864438057,
0.017982907593250275,
0.017530981451272964,
-0.024824976921081543,
0.029987744987010956,
0.01132055651396513,
-0.027110129594802856,
-0.00694984570145607,
0.0033185561187565327,
0.01675398275256157,
-0.02027168497443199,
-... |
angular-cli/README.md_0_3860 | [-0.017614876851439476,-0.01029879879206419,-0.055440571159124374,-0.024731185287237167,0.0256740804(...TRUNCATED) |
angular-cli/README.md_3860_10863 | [-0.018557939678430557,0.000013803519323118962,-0.0550418458878994,0.012307126075029373,-0.001002177(...TRUNCATED) |
angular-cli/constants.bzl_0_1097 | [-0.015817172825336456,-0.05315708369016647,-0.02162054181098938,0.014346320182085037,0.010836268775(...TRUNCATED) |
angular-cli/CONTRIBUTING.md_0_7325 | [-0.009504705667495728,0.008592264726758003,-0.02597563900053501,-0.033077992498874664,-0.0104032615(...TRUNCATED) |
angular-cli/CONTRIBUTING.md_7325_13061 | [-0.013979606330394745,-0.00517691858112812,-0.034649308770895004,-0.026897329837083817,-0.004545263(...TRUNCATED) |
angular-cli/BUILD.bazel_0_1583 | [-0.0019623476546257734,-0.02560703083872795,-0.019227199256420135,-0.012076658196747303,0.022627346(...TRUNCATED) |
End of preview. Expand in Data Studio
FreshStack Leaderboard: Commercial Embeddings
Precomputed document and query embeddings from commercial (API-only) embedding models on the
FreshStack benchmark. These are the vectors behind the scores and run files in
freshstack/leaderboard-results.
You can use them to reproduce retrieval runs without calling the paid API again.
- Corpus:
freshstack/corpus-oct-2024 - Queries:
freshstack/queries-oct-2024(testsplit)
Models
| Model | Provider | Dim | Query input type | Document input type | Max tokens (q / d) |
|---|---|---|---|---|---|
embed-v5.0-fast |
Cohere API v2 /embed |
2048 | search_query |
search_document |
2048 / 2048 |
embed-v5.0-pro |
Cohere API v2 /embed |
2048 | search_query |
search_document |
2048 / 2048 |
Document text is title + ' ' + text, truncated at the END. The embeddings are float32 and L2-normalized, so
use cosine similarity or dot product to score them. Each model folder has a configuration.json with the full
run settings.
Layout
<model>/configuration.json
<model>/<domain>/corpus.parquet # columns: _id (string), embedding (float32[2048])
<model>/<domain>/queries.parquet # columns: _id (string), embedding (float32[2048])
Each <model>-<domain> config has two splits, corpus and queries.
| Model | Domain | Corpus | Queries | Dim |
|---|---|---|---|---|
embed-v5.0-fast |
angular | 117,286 | 129 | 2048 |
embed-v5.0-fast |
godot | 25,477 | 99 | 2048 |
embed-v5.0-fast |
langchain | 49,505 | 203 | 2048 |
embed-v5.0-fast |
laravel | 52,341 | 184 | 2048 |
embed-v5.0-fast |
yolo | 27,207 | 57 | 2048 |
embed-v5.0-pro |
angular | 117,286 | 129 | 2048 |
embed-v5.0-pro |
godot | 25,477 | 99 | 2048 |
embed-v5.0-pro |
langchain | 49,505 | 203 | 2048 |
embed-v5.0-pro |
laravel | 52,341 | 184 | 2048 |
embed-v5.0-pro |
yolo | 27,207 | 57 | 2048 |
Usage
import numpy as np
from datasets import load_dataset
repo = "freshstack/leaderboard-commercial-embeddings"
corpus = load_dataset(repo, "embed-v5.0-pro-angular", split="corpus").with_format("numpy")
queries = load_dataset(repo, "embed-v5.0-pro-angular", split="queries").with_format("numpy")
scores = queries["embedding"] @ corpus["embedding"].T # cosine similarity (vectors are normalized)
top100 = np.argsort(-scores, axis=1)[:, :100]
You can also read the files directly, for example with pd.read_parquet("hf://datasets/{repo}/embed-v5.0-pro/angular/corpus.parquet").
License
The embeddings are derived from the FreshStack corpus and queries, which are licensed under CC-BY-SA-4.0. The vectors were generated with Cohere's API, so any use of them is also subject to Cohere's terms of use.
- Downloads last month
- 120