"""README.md / SELECTION.md of the MTEB-lite release (used by mteb_lite_release.py).""" from __future__ import annotations import numpy as np SUITE_NAME = {"eng": "MTEB(eng, v2) (41 tasks)", "ara": "Arabic MMTEB (42 test tasks)"} def _f(x, nd=3): return "-" if x is None or (isinstance(x, float) and np.isnan(x)) else f"{x:.{nd}f}" def _agg_table(rep: dict) -> str: combo = rep["choice"]["combo"] lines = ["| protocol | estimate | macro tau | macro tau (resolvable pairs) | macro Pearson | macro MAE | mean tau | mean MAE | held-out models |", "|---|---|---|---|---|---|---|---|---|"] for proto, label in (("family", "family-level LOMO"), ("distinct", "distinct-model LOMO")): A = rep["aggregate"].get(proto, {}) for key, name in ((combo, "**MTEB-lite**"), ("random", "random K (single draw)"), ("pool", "whole candidate pool")): v = A.get(key) if not v: continue lines.append(f"| {label} | {name} | {_f(v['macro']['kendall'])} | {_f(v['macro'].get('tau_resolvable'))} | " f"{_f(v['macro']['pearson'])} | {_f(v['macro']['mae'], 2)} | {_f(v['mean']['kendall'])} | " f"{_f(v['mean']['mae'], 2)} | {v['macro']['n']} |") return "\n".join(lines) def _suite(cfg: str) -> str: return "ara" if cfg.startswith("ara") else "eng" def readme(st: dict, reps: dict, mans: dict, models: list[dict], cal: dict, notes: dict) -> str: cfgs = list(mans) suites = sorted({_suite(c) for c in cfgs}) model_suites = [su for su in ("eng", "ara") if f"tasks_with_full_score_{su}" in models[0]] size_rows = [] for cfg in cfgs: s = st.get(cfg, {}) m = mans[cfg] size_rows.append(f"| `{cfg}` | {m['benchmark']} | {m['n_tasks']} | {m['K']} | {m['selection_version']} | " f"{s.get('items', '-')} | {s.get('texts', '-')} | {_f(s['tokens'] / 1e6, 2) if s else '-'}M | " f"~{_f(s.get('minutes_0.6B'), 1) if s else '-'} min |") val = "\n\n".join(f"### `{cfg}` (K = {mans[cfg]['K']}, selection {mans[cfg]['selection_version']}; chosen variant " f"{reps[cfg]['choice']['combo']})\n\n{_agg_table(reps[cfg])}" for cfg in cfgs) cal_txt = [] for cfg in cfgs: rows = cal[cfg] r = [x["fit_pearson_r"] for x in rows if x["fit_pearson_r"] == x["fit_pearson_r"]] sd = [x["fit_residual_sd"] for x in rows] cal_txt.append(f"| `{cfg}` | [`{cfg}/calibration.csv`]({cfg}/calibration.csv) | {len(rows)} | " f"{_f(float(np.median(r)))} | {_f(float(np.min(r)))} | {_f(100 * float(np.median(sd)), 2)} |") weak = [] for cfg in cfgs: w = [f"{x['task']} (r={x['fit_pearson_r']:.2f})" for x in cal[cfg] if not (x["fit_pearson_r"] >= 0.3)] if w: weak.append(f"`{cfg}`: " + ", ".join(w)) weak_txt = ("* **Weakly calibrated tasks** (in-sample fit r < 0.3; their estimates stay close to the reference-model mean, " "so they add little ranking signal). A constant subset statistic, e.g. 20 clustering items over hundreds of " "labels where every item is its own cluster, gives r = 0:\n " + "; ".join(weak)) if weak else "" role_cols = [c for c in models[0] if c.startswith("role_")] mrows = [] for x in sorted(models, key=lambda x: ({"ADIP release": 0, "Qwen base": 1, "public": 2, "ADIP merge": 3}[x["type"]], x["display_name"])): used = [c[5:] for c in role_cols if x[c].startswith("selection")] full = ", ".join(f"{su}: {x[f'tasks_with_full_score_{su}']}" for su in model_suites if x.get(f"tasks_with_full_score_{su}") not in (None, "", "0", 0)) mrows.append(f"| {x['display_name']} | `{x['internal_id']}` | {x['type']} | {x['suites_scored']} | {full or '-'} | " f"{', '.join(used) if used else 'not used'} |") ara_txt = notes.get("ara", "") eng_note = notes.get("eng", "") return f"""--- license: other license_name: per-task-source-licenses pretty_name: MTEB-lite viewer: false tags: - mteb - embeddings - benchmark - evaluation language: - en{chr(10) + '- ar' if 'ara' in suites else ''} --- # MTEB-lite **MTEB-lite** is a set of small subsets of embedding benchmarks that estimate a text-embedding model's full-benchmark scores at a small fraction of the cost: {chr(10).join(f'* `{c}`: {SUITE_NAME[_suite(c)]}, K = {mans[c]["K"]} items per task' for c in cfgs)} For every task, a fixed set of items (queries, test examples, sentence pairs, clustering items, documents, bitext pairs) was chosen from a candidate pool. The subset's statistic, after a per-task linear calibration, tracks the full-task main score across a diverse set of reference models. This repository contains **only item ids, selection weights, item-model parameters and calibration coefficients, plus code**. No text from any dataset is redistributed. `mteb_lite.py` rebuilds the selected items from the original datasets through `mteb` (dataset revisions pinned by mteb {mans[cfgs[0]]['mteb_version']}) and evaluates a model with mteb's own per-task prompts. ## Configs | config | benchmark | tasks | K (items / task) | selection | items | texts to encode | tokens | encode time, 0.6B model* | |---|---|---|---|---|---|---|---|---| {chr(10).join(size_rows)} \\* Qwen3-Embedding-0.6B at about 30k tokens/s on one GPU, the rate we measured in our full MTEB runs. Tokens are counted with the Qwen3 tokenizer, including instruction prefixes{'; Arabic texts are capped at 512 tokens, as in our Arabic MMTEB eval' if 'ara' in suites else ''}. {eng_note} Layout: `/manifest.json` holds the task list, main scores, dataset paths / revisions, K, estimators, calibration and reference models. `/.json` holds the selected ids plus whatever scoring needs, and `/calibration.csv` the per-task calibration. `mteb_lite.py` is the loader + evaluator, `validation/.{{md,json}}` the validation reports, `SELECTION.md` the selection method, `models.csv` the reference models, and `code/` the scripts used to build pools, score models, select and evaluate. ## Usage ```bash pip install "mteb==1.39.7" sentence-transformers numpy scikit-learn python mteb_lite.py --model sentence-transformers/all-MiniLM-L6-v2 --config eng-super --out result.json ``` ```python from mteb_lite import evaluate, load_task res = evaluate("BAAI/bge-small-en-v1.5", config="eng") # an mteb model name, or any mteb Encoder object res["estimated_macro"], res["estimated_mean"] # estimated benchmark macro (mean over task types) / mean (x100) res["rows"]["ArguAna"] # subset statistic + calibrated full-task estimate task = load_task("eng", "STS12") # the selected items with their texts ``` The model is called exactly as mteb calls it: `model.encode(texts, task_name=..., prompt_type=query|document|None)`. Instruction models therefore get their own per-task instructions. Pass `mteb.get_model(name)` (or just the name) for models mteb knows; wrap other models in an mteb Encoder. For the Arabic configs, set the model's max sequence length to 512 to match the reference runs. ## Calibration For every task, `full_estimate = a + b * subset_statistic`. The subset statistic is one of: * `weighted_mean`: the anchor-weighted mean of the per-item outcomes; * `pirt`: the p-IRT estimate of the pool score (observed items + item-model predictions for the rest); * `gp`: the average of the two; * `set_metric`: Spearman / weighted AP / V-measure of the selected items (STS, pair classification, clustering). `a` and `b` are least-squares fits over the config's reference models that have a full score for the task. The benchmark estimates are the **macro** (mean over task types of the mean over tasks) and the **mean** (over tasks) of the per-task estimates, x100. The coefficients are in every `/.json`, in `/manifest.json`, and in: | config | file | tasks | median fit r | min fit r | median residual SD (points) | |---|---|---|---|---|---| {chr(10).join(cal_txt)} The fit r and residual SD are in-sample (on the fitting models). Out-of-sample accuracy is in the validation tables below. ## What each task file contains (ids only) * **Retrieval:** the selected query ids, and per query its candidate document ids (BM25 top-30 over the full corpus + its relevant documents, at most 100) and its qrels; also the ids of the random corpus documents shared by all queries. A query is ranked against its own candidates + the shared random documents, and its nDCG@10 is the item outcome. * **Reranking:** the selected samples (row indices). Each keeps its own candidate list from the dataset. The outcome is AP (nDCG@10 for MIRACL-style main scores). * **Classification:** the selected test rows, the train rows that mteb's undersampling protocol draws (n_experiments x samples_per_label per label, seed 42) and the per-experiment row lists. The outcome is the fraction of the logistic-regression experiments that are correct. * **BitextMining:** the selected gold pairs plus a fixed set of distractor target sentences per language pair. The outcome is 1 if the gold target is the nearest neighbour. * **STS / PairClassification / Clustering:** the selected pair / item rows (+ pair label weights, clustering label depth), scored as sets. * **Summarization:** the selected document ids; the outcome is the per-document Spearman. * **p-IRT tasks:** additionally the ids, weights and item parameters of the whole candidate pool (no texts). **Id scheme.** We use the source dataset's own ids where they exist (retrieval query / document ids, SummEval document ids). Otherwise an id is ``, the row index in the evaluation split of the dataset exactly as `mteb==1.39.7`'s `task.load_data()` returns it (after mteb's own `dataset_transform`). The prefix is `r` (reranking sample), `t` (test example), `p` (pair, flattened pair list), `s` (STS pair), `c` (clustering item) or `b` (bitext gold pair). Classification train rows are train-split row indices. Retrieval task files also store a 16-hex SHA-1 of every needed query / document text, and the loader matches texts by these hashes. This is needed because some mteb loaders (e.g. MintakaRetrieval, MKQARetrieval, PublicHealthQA, XPQARetrieval) number queries and documents via `enumerate(set(...))`, so their ids change from process to process. The hashes also act as an integrity check. In multi-subset tasks (the Arabic configs), ids are prefixed with `|`, and bitext targets are `|`. Each task file records the dataset path, revision, eval split and hf_subset(s). {ara_txt} ## Reference models Selection and calibration use a diverse **reference set**: our model families, each thinned to at most 3 models (best / median / worst by full score), plus every public and base model. Every model that was scored is listed below and in `models.csv`, which also gives each model's role per config, internal id, family and merge recipe. Our merges are described only by their merge recipe. | model | internal id | type | suites scored | tasks with full scores | used in configs | |---|---|---|---|---|---| {chr(10).join(mrows)} ## Validation Validation is leave-one-out, with selection and calibration redone without the held-out models: * **family-level LOMO** holds out a whole model family at a time. This is the honest test of generalisation to a new kind of model. * **distinct-model LOMO** holds out one reference model at a time. We report the Kendall tau / Pearson / MAE between the estimated and the true benchmark macro and mean (x100) for the held-out models that have full scores on all tasks. "Resolvable pairs" are model pairs whose true scores differ by at least 0.5 points. The baselines are K uniformly random items (metrics averaged over 10 single draws) and the whole candidate pool, which is the ceiling for any subset of it. Per-task tables are in `validation/`. {val} ## Limitations * **The reference set is dominated by Qwen3-Embedding-based models:** our merges of fine-tuned Qwen3-Embedding-0.6B and the Qwen base models, plus the public models listed above. The calibration is linear, so estimates for very different models are extrapolations. Expect the ranking to be more reliable than the absolute scores. * **Set-level tasks** (STS Spearman, pair-classification AP, clustering V-measure) are not sums over items, so per-item IRT does not apply. They are scored as sets on anchor points of per-item outcomes. Clustering with 20 items is very noisy. {weak_txt} * **Pool ceiling.** Items come from candidate pools: up to 1,000 queries per retrieval task with BM25-30 candidates + random documents, and up to 1,000 items per other task. A retrieval query is scored against its candidate set, not the full corpus. Even the whole pool only reaches the tau in the "whole candidate pool" rows. Much of the residual tau error comes from near-tied models (less than 0.5 points apart); see the resolvable-pairs column. * **Main-score families.** Pair-classification main scores are max-AP over similarity functions; the subsets use cosine AP and rely on calibration. Bitext mining uses nearest-neighbour accuracy against fixed distractors in place of mteb's F1. ## License This repository ships only ids, weights, coefficients and code. Each task's data remains under its source dataset's license, and `mteb_lite.py` downloads it from the original source at evaluation time. Code: Apache-2.0. ## Citation ```bibtex @misc{{mteb_lite_2026, title = {{MTEB-lite: validated small subsets of MTEB(eng, v2) and Arabic MMTEB}}, author = {{ADIP team, QCRI}}, year = {{2026}}, note = {{https://huggingface.co/datasets/ADIP-QCRI/MTEB-lite}} }} ``` Please also cite MTEB / MMTEB and the source datasets of the tasks you report. ## Contact easgari@hbku.edu.qa """ def selection_md(reps: dict, mans: dict) -> str: fam_txt = [] for cfg, r in reps.items(): fams = r.get("families", {}) fam_txt.append(f"**`{cfg}`** ({len(r['reference'])} models, {len(fams)} families):\n\n" + "\n".join(f"* {k}: kept {', '.join(v['kept'])} (of {len(v['members'])})" for k, v in sorted(fams.items()))) chosen = "\n".join(f"* `{cfg}` (selection {mans[cfg]['selection_version']}): per-item tasks **{rep['choice']['per_item']}**, " f"set-metric tasks **{rep['choice']['set_metric']}**, K = {mans[cfg]['K']}" for cfg, rep in reps.items()) return f"""# 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. {chr(10).join(fam_txt)} ## 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 {chosen} 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. """