TIPS-English

Frozen lite English embedding-evaluation dataset — the downsampled MTEB(eng, v1) + bitext subsets used by the tips-evaluation framework (english/ runner, --lite), published so evaluation boxes download this (~0.6 GB) instead of the ~122 GB original corpora. The English counterpart of TIPS-Korean.

Built from the 56 MTEB(eng, v1) tasks + BUCC/Tatoeba with seeded, model-independent downsampling (english/downsample.py, seed 42):

Task type	Reduction
Retrieval / Reranking	500 queries, corpus capped at 10,000 (positives always kept; reranking keeps the kept queries' candidate lists)
Classification	2,048 rows per eval split + 10,000-row train pool
Clustering (legacy)	4 experiments × ≤10,000 docs
STS / PairClassification	2,048 pairs
BUCC	500 gold pairs, 10,000-candidate pool (gold indices remapped)
SummEval, Tatoeba	full size (already small)

Scores on this dataset are comparable across models, NOT to full MTEB(eng, v1) or the public leaderboard (smaller corpora inflate retrieval absolute numbers).

Layout

One folder per mteb task (manifest.json lists all 69 leaf tasks, the exact lite config, mteb version, and per-task content hashes):

<Task>/<subset>/<split>/queries.jsonl      retrieval/reranking: id, text
<Task>/<subset>/<split>/corpus.jsonl       id, title, text
<Task>/<subset>/<split>/qrels.jsonl        query-id, corpus-id, score
<Task>/<subset>/<split>/top_ranked.jsonl   query-id, corpus-ids (reranking)
<Task>/[<subset>/]<split>.jsonl            all other task types (task rows)
<Task>/[<subset>/]<split>.dict.json        dict-shaped leaves (BUCC)
prompts.json                               tuned per-task instructions
manifest.json                              provenance + hashes

• prompts.json — per-task query instructions selected by a documented grid search (english/sweep_instructions.py --lite) with Qwen/Qwen3-VL-Embedding-8B; applied automatically by the runner (override with --task-prompts).
• filtered_out_vl (MindSmallReranking queries.jsonl) — flags 13/500 queries (per-query AP ≤ 0.015 under Qwen3-VL-Embedding-8B). This is a MODEL-DEPENDENT filter, disclosed in manifest.json; rows are kept in the data and only dropped when the runner is invoked with --apply-vl-filter (default off = honest full set) — the same convention as TIPS-Korean's column of the same name.

Usage

# tips-evaluation repo — frozen dataset is the default lite source:
LITE=1 OFFLINE=0 bash scripts/eval.sh <hf-model-id> english
VL_FILTER=1 LITE=1 bash scripts/eval.sh <hf-model-id> english  # cert subset

# or directly:
python -m english.run_english --model <hf-model-id> --lite \
    --lite-dataset Cartinoe5930/TIPS-English [--apply-vl-filter]

Regenerate / verify from the original sources (hash-checked round trip):

python -m english.freeze --output-dir eval_data/tips-english \
    --prompts lite_prompts.json --vl-filter vl_filter_english.json --verify

Sources: the mteb/* HuggingFace datasets pinned by mteb 2.18.0 task metadata (revisions recorded per task in mteb). For research/evaluation use only; source datasets keep their original licenses.