Skip to content

2.27.15

Choose a tag to compare

@KennethEnevoldsen KennethEnevoldsen released this 17 Nov 15:35
· 16 commits to main since this release

2.27.15 (2026-11-17)

Chore

  • chore: update the citation cache [skip ci] (926745a)

Documentation

  • docs: fix baseline-random-encoder model name in adding_a_dataset (#5898)

Signed-off-by: tonydzi <dzyatkovskiy.a@gmail.com>
Co-authored-by: tonydzi <dzyatkovskiy.a@gmail.com> (a0444b1)

  • docs: fix image paths in the vLLM wrapper guide and an installation anchor (#5886) (1f61c3f)

Fix

  • fix: keep run_settings of cached subsets when resuming an evaluation (#5906)

  • fix: keep run_settings of cached subsets when resuming an evaluation

save_to_cache recorded run settings for every subset in the result, including
subsets loaded from the cache, so a resumed run replaced their encode_kwargs
and package versions. Only the subsets evaluated in the current run are
recorded now, and the intermediate save passes encode_kwargs.

  • update

Co-authored-by: Roman Solomatin <36135455+Samoed@users.noreply.github.com> (38e934a)

Refactor

  • refactor: move TurHistQuad, GeorgianFAQ, NorQuad and SadeemQuestion retrieval to v2 dataset format (#5891) (1b47bb5)

Unknown

  • model: Add mzahran1/arabic-english-bge-m3-retrieval, e5-base-retrieval and e5-small-retrieval-v2 (#5899)

  • model: add mzahran1 Arabic-English bge-m3, e5-base and e5-small v2 retrieval models

Signed-off-by: Mahmoud Zahran <200594164+mahmoudzahran01@users.noreply.github.com>

  • model: shorten comments in mzahran_models

Signed-off-by: Mahmoud Zahran <200594164+mahmoudzahran01@users.noreply.github.com>


Signed-off-by: Mahmoud Zahran <200594164+mahmoudzahran01@users.noreply.github.com>
Co-authored-by: Mahmoud Zahran <200594164+mahmoudzahran01@users.noreply.github.com> (fc3cc8a)

  • model: add f4roukb/denserepr-code-48M-pretrain-EN-v1 (#5893)

  • model: add f4roukb/denserepr-code-48M-pretrain-EN-v1

Dense code-retrieval embedding model (47.7M parameters, 384-d, 8,192 tokens), continued contrastive
pretraining of ibm-granite/granite-embedding-small-english-r2. Trained on the CoIR train splits, listed in
training_datasets. Loaded in bf16, as evaluated.

  • Delete mteb_mock_run_results.md

Co-authored-by: Roman Solomatin <samoed.roman@gmail.com> (c090baa)

  • model / fix: declare training datasets for embed-v1 / asym-v1; fix v3 training data (#5890)

model: declare training datasets for opensearch embed-v1 / sparse asym-v1, fix v3 task name (cb4e764)