Hugging Face's logo

Datasets:
ADIP-QCRI
/
MTEB-lite

Languages:
English
Arabic
Tags:
mteb
embeddings
benchmark
evaluation
License:
Dataset card Files
xet
 Community
MTEB-lite / SELECTION.md
ehsaneddin's picture
ehsaneddin
Add ara-super (Arabic MMTEB, K=20): selection, calibration, validation
f8f07da verified
preview code
|
Raw History Blame Contribute Delete
8.78 kB

MTEB-lite: selection method

1. Candidate pools (per task, from the eval split mteb uses)

  • English (eng*): MTEB(eng, v2), with each task's English subset and test split.
  • Arabic (ara*): the 42 Arabic MMTEB test tasks. Each is restricted to its Arabic hf_subsets (Arabic, and Arabic-English pairs) and to the split our Arabic MMTEB evaluation uses. Multi-subset tasks get their items split evenly over subsets, and their task score is the mean over subsets, as in mteb. Texts are capped at 512 tokens.

Per task type:

  • Retrieval: up to 1,000 queries with at least one relevant document (seed 0). Each query keeps its own candidate list: the BM25 top 30 over the full corpus (Okapi k1=0.9, b=0.4) plus its relevant documents (at most 100). Random corpus documents shared by all queries are added: 1,000, or max(200, 1000/#subsets) per subset for multi-subset tasks.
  • Reranking: up to 1,000 samples with their own candidate lists. Arabic MIRACLReranking was cut to 300 queries because of encode cost.
  • Classification: up to 1,000 test examples (label-stratified). The train rows are the ones mteb's undersampling protocol draws (seed 42).
  • Pair classification: up to 1,000 pairs, balanced by label. AP is reweighted to the full label prior.
  • STS: up to 1,000 pairs, stratified by 5 gold-score bins.
  • Clustering: up to 1,000 items, stratified by label; hierarchical label levels are kept.
  • Summarization: all SummEval documents.
  • Bitext mining (Arabic): up to 1,000 gold pairs. Each language pair gets a fixed set of up to 1,000 distractor targets, and parallel datasets reuse the same rows.

2. Per-item outcomes (one model x one item)

type outcome
Retrieval nDCG@10 of the query
Reranking average precision (nDCG@10 for MIRACL-style main scores)
Classification fraction of the n_experiments logistic regressions (mteb protocol) that are correct
Summarization Spearman(human relevance, max cosine to references) of the document
Bitext mining 1 if the gold target is the nearest of target + distractors
STS 1 - abs(percentile(cos) - percentile(gold))
Pair classification correctness at the accuracy-optimal cosine threshold
Clustering 1 if the item's k-means cluster majority label equals its label

3. Reference models and full scores

Model families are merge recipes or training runs; every public or base model is its own family. At most 3 models are kept per family (best / median / worst by full score). Full scores come from our own mteb 1.39.7 runs for our models and the Qwen base models. For public models they come from the official MTEB results (English: test split of MTEB(eng, v2); Arabic: the same Arabic subsets / split, available for only 15-31 of the 42 tasks per public model; a model enters a task only where it has a full score). For English, 18 additional public models (BGE, E5, GTE, Arctic-Embed, mxbai, MiniLM / mpnet, LaBSE, GIST, UAE) were scored in selection v3 to add diversity.

eng (28 models, 15 families):

  • A-v3@6750: kept A-lin0.6, A-lin0.5, A-ties-d0.1-l0.5 (of 5)

  • AS3: kept AS3-dareties-p0.9-l1.0, AS3-ties-d0.3-l1.0, AS3-ties-d0.3-l0.5 (of 4)

  • AV2: kept AV2-dareties-p0.9-l1.0, AV2-lin1.0, AV2-lin0.5 (of 6)

  • W: kept W-A0.7-S3-0.3-l0.5 (of 1)

  • all-MiniLM-L6-v2: kept all-MiniLM-L6-v2 (of 1)

  • bge-m3: kept bge-m3 (of 1)

  • e5-base-v2: kept e5-base-v2 (of 1)

  • embeddinggemma-300m: kept embeddinggemma-300m (of 1)

  • multilingual-e5-base: kept multilingual-e5-base (of 1)

  • multilingual-e5-small: kept multilingual-e5-small (of 1)

  • qwen3-0.6b: kept qwen3-0.6b, qwen3-0.6b+norm (of 2)

  • qwen3-4b: kept qwen3-4b (of 1)

  • v1-F/P: kept F-lin1.0, F-lin0.5, F-ties-d0.1-l0.5 (of 7)

  • v4: kept V-lin0.7, S-lin0.5, V-ties-d0.1-l0.5 (of 5)

  • v5: kept S5-lin0.5, V5-ties-d0.3-l0.7, V5-ties-d0.1-l0.5 (of 5) eng-super (46 models, 33 families):

  • A-v3@6750: kept A-lin0.6, A-lin0.5, A-ties-d0.1-l0.5 (of 5)

  • AS3: kept AS3-dareties-p0.9-l1.0, AS3-ties-d0.3-l1.0, AS3-ties-d0.3-l0.5 (of 4)

  • AV2: kept AV2-dareties-p0.9-l1.0, AV2-lin1.0, AV2-lin0.5 (of 6)

  • GIST-small-Embedding-v0: kept GIST-small-Embedding-v0 (of 1)

  • LaBSE: kept LaBSE (of 1)

  • UAE-Large-V1: kept UAE-Large-V1 (of 1)

  • W: kept W-A0.7-S3-0.3-l0.5 (of 1)

  • all-MiniLM-L12-v2: kept all-MiniLM-L12-v2 (of 1)

  • all-MiniLM-L6-v2: kept all-MiniLM-L6-v2 (of 1)

  • all-mpnet-base-v2: kept all-mpnet-base-v2 (of 1)

  • arctic-embed-l: kept arctic-embed-l (of 1)

  • arctic-embed-m-v1.5: kept arctic-embed-m-v1.5 (of 1)

  • arctic-embed-xs: kept arctic-embed-xs (of 1)

  • bge-base-en-v1.5: kept bge-base-en-v1.5 (of 1)

  • bge-large-en-v1.5: kept bge-large-en-v1.5 (of 1)

  • bge-m3: kept bge-m3 (of 1)

  • bge-small-en-v1.5: kept bge-small-en-v1.5 (of 1)

  • e5-base-v2: kept e5-base-v2 (of 1)

  • e5-large-v2: kept e5-large-v2 (of 1)

  • e5-small-v2: kept e5-small-v2 (of 1)

  • embeddinggemma-300m: kept embeddinggemma-300m (of 1)

  • gte-large: kept gte-large (of 1)

  • gte-small: kept gte-small (of 1)

  • multilingual-e5-base: kept multilingual-e5-base (of 1)

  • multilingual-e5-large: kept multilingual-e5-large (of 1)

  • multilingual-e5-small: kept multilingual-e5-small (of 1)

  • mxbai-embed-large-v1: kept mxbai-embed-large-v1 (of 1)

  • paraphrase-multilingual-mpnet-base-v2: kept paraphrase-multilingual-mpnet-base-v2 (of 1)

  • qwen3-0.6b: kept qwen3-0.6b, qwen3-0.6b+norm (of 2)

  • qwen3-4b: kept qwen3-4b (of 1)

  • v1-F/P: kept F-lin1.0, F-lin0.5, F-ties-d0.1-l0.5 (of 7)

  • v4: kept V-lin0.7, S-lin0.5, V-ties-d0.1-l0.5 (of 5)

  • v5: kept S5-lin0.5, V5-ties-d0.3-l0.7, V5-ties-d0.1-l0.5 (of 5) ara-super (39 models, 21 families):

  • A-v3@6750: kept A-dare-p0.9-l1.0, A-lin0.6, A-lin0.5 (of 3)

  • AS3: kept AS3-dareties-p0.9-l1.0, AS3-ties-d0.3-l1.0, AS3-ties-d0.3-l0.5 (of 3)

  • AV2: kept AV2-dareties-p0.9-l1.0, AV2-lin1.0, AV2-lin0.5 (of 3)

  • FV: kept FV-ties-d0.1-l0.5, FV-ties-d0.3-l0.7 (of 2)

  • LaBSE: kept LaBSE (of 1)

  • Vq: kept Vq-lin0.6, Vq-lin0.4, Vq-lin0.7 (of 3)

  • W: kept W-A0.7-V2-0.3-l0.5, W-F0.7-V0.3-l0.7, W-A0.7-S3-0.3-l0.5 (of 3)

  • all-MiniLM-L6-v2: kept all-MiniLM-L6-v2 (of 1)

  • bge-m3: kept bge-m3 (of 1)

  • e5-base-v2: kept e5-base-v2 (of 1)

  • embeddinggemma-300m: kept embeddinggemma-300m (of 1)

  • multilingual-e5-base: kept multilingual-e5-base (of 1)

  • multilingual-e5-large: kept multilingual-e5-large (of 1)

  • multilingual-e5-small: kept multilingual-e5-small (of 1)

  • paraphrase-multilingual-MiniLM-L12-v2: kept paraphrase-multilingual-MiniLM-L12-v2 (of 1)

  • paraphrase-multilingual-mpnet-base-v2: kept paraphrase-multilingual-mpnet-base-v2 (of 1)

  • qwen3-0.6b: kept qwen3-0.6b, qwen3-0.6b+norm (of 2)

  • qwen3-4b: kept qwen3-4b (of 1)

  • v1-F/P: kept F-lin1.0, F-lin0.4, F-lin0.5 (of 3)

  • v4: kept V-lin1.0, V-lin0.5, V-lin0.1 (of 3)

  • v5: kept V5-lin1.0, S5-lin0.5, V5-lin0.7 (of 3)

4. Item models and selection

The model ability is theta = the full-task score, standardised over the training models. Per item:

  • continuous outcomes: a linear-Gaussian factor model, y = a * theta + b + e;
  • binary outcomes: 2PL, sigmoid(alpha * theta + beta), L2-regularised.

Candidates evaluated:

  • anchor: tinyBenchmarks anchor points. K-means groups the items' outcome vectors over the training models; the anchor is the item nearest to each centroid, and its weight is the cluster's share of the pool.
  • fisher: a greedy maximum of sum over models of log(eps + information).
  • estimators: the weighted subset mean; p-IRT (MAP theta from the selected items, with the rest of the pool predicted); and gp, their average.
  • set-metric tasks: stratified random, or anchor points on the per-item outcomes. Each estimate is mapped by a per-task linear calibration fitted on the reference models.

5. Validation protocol

  • family-level LOMO: for every family, select + calibrate without it and predict its models.
  • distinct-model LOMO: the same, one reference model at a time.
  • metrics: Kendall tau / Pearson / MAE of the predicted vs true benchmark macro (mean over type means) and mean (over tasks), for the held-out models with all full scores. We also report tau on resolvable pairs (difference of at least 0.5 points).
  • baselines: uniform random K (metrics averaged over 10 single draws) and the whole pool.

6. Chosen variants

  • eng (selection v2): per-item tasks anchor/mean, set-metric tasks anchorset, K = 100
  • eng-super (selection v3): per-item tasks fisher/mean, set-metric tasks anchorset, K = 20
  • ara-super (selection v4): per-item tasks anchor/gp, set-metric tasks strat, K = 20

Variants were chosen among {anchor, fisher} x {mean, p-IRT, gp} x {stratified, anchor set} on the same LOMO folds, which makes the numbers mildly optimistic.