alibayram/embeddingmagibu2-distill-mix

embeddingmagibu2 Distillation Mix

A Turkish multi-task text mix with precomputed teacher vectors from google/embeddinggemma-2, for offline distillation of alibayram/embeddingmagibu2 on every task the teacher handles: retrieval, question answering, semantic similarity, NLI, classification, clustering and Turkish–English bitext. Wikipedia rows in 40 languages are mixed in so the student keeps its other languages.

Status: splits are pushed as soon as they are encoded (holdout → pct01 → pct10 → pct30 → rest). Training can start on pct01 while the larger splits are still being encoded.

Splits

Every source is shuffled once and cut into the same proportions, so every split contains every task.

Split	Content	Rows
holdout	50 rows per source	1,150
pct01	1% of every source	7,006
pct10	10% of every source	70,175
pct30	30% of every source	210,525
rest	the remaining 59%	414,044

Columns

Column	Description
source	source dataset (and config) of the row
a	anchor text (query, question, sentence, or the single text of classification/clustering rows)
b	positive / paired text, empty if the row has a single text
neg	hard negative from the source, empty if none
a_emb, b_emb, neg_emb	teacher vectors (768-d, L2-normalized, float32); null where the text is empty

Texts are stored exactly as embedded: the teacher's task prompt followed by the text clipped to 2000 characters, e.g. task: search result | query: ..., title: none | text: ..., task: sentence similarity | query: .... Feed the same strings to the student. Labels of the sources (sentiment, category, similarity scores) are not kept: the teacher's vectors are the targets.

Sources

Rows per split: holdout / pct01 / pct10 / pct30 / rest.

Source	Prompts	holdout	pct01	pct10	pct30	rest
F2LLM-v2/webfaq_tur	query → document	50	799	7,995	23,985	47,171
F2LLM-v2/mqa_tr	query → document	50	399	3,995	11,985	23,571
turkish-extractive-qa-1.5m	QuestionAnswering → document	50	599	5,995	17,985	35,371
turkish-retrieval-constructed-1m	query → document	50	99	995	2,985	5,871
TQuad-2	QuestionAnswering → document	50	149	1,495	4,485	8,821
cleaned_turkish_embedding_model_training_data_colab/doktorsitesi_question_answer_pairs	query → document	50	399	3,995	11,985	23,571
cleaned_turkish_embedding_model_training_data_colab/mfaq_question_answer_pairs	query → document	50	299	2,995	8,985	17,671
cleaned_turkish_embedding_model_training_data_colab/legal_nli_TR_V1_title_text_pairs	query → document	50	299	2,995	8,985	17,671
cleaned_turkish_embedding_model_training_data_colab/babel-briefings_question_context_pairs	query → document	50	299	2,995	8,985	17,671
cleaned_turkish_embedding_model_training_data_colab/OpenOrca-tr_question_answer_pairs	query → document	50	299	2,995	8,985	17,671
cleaned_turkish_embedding_model_training_data_colab/InstrucTurca_question_answer_pairs_0	query → document	50	299	2,995	8,985	17,671
cleaned_turkish_embedding_model_training_data_colab/mlsum_title_text_pairs	query → document	50	249	2,495	7,485	14,721
cleaned_turkish_embedding_model_training_data_colab/ccnews_2019_pairs	query → document	50	249	2,495	7,485	14,721
cleaned_turkish_embedding_model_training_data_colab/TDK_Sozluk-Turkish_title_text_pairs	query → document	50	149	1,495	4,485	8,821
all-nli-tr/triplet	STS → STS	50	799	7,995	23,985	47,171
F2LLM-v2/xnli_tr	PairClassification → PairClassification	50	199	1,995	5,985	11,771
stsb-tr	STS → STS	50	79	795	2,385	4,691
turkish-sentiment-analysis-dataset	Classification	50	249	2,495	7,485	14,721
turkish-financial-sentiment-256k	Classification	50	149	1,495	4,485	8,821
ttc4900	Clustering	50	48	485	1,455	2,862
parallel-sentences-talks/en-tr	BitextMining → BitextMining	50	299	2,995	8,985	17,671
parallel-sentences-tatoeba/en-tr	BitextMining → BitextMining	50	199	1,995	5,985	11,771
wikipedia-40-langs-with-embeddings-embeddinggemma2	query → document	50	399	3,995	11,985	23,571

How it was built

• Teacher: google/embeddinggemma-2 in bfloat16 via sentence-transformers; vectors are normalized.
• All splits of each source were merged, shuffled with seed 0 and the first N rows were taken.
• Texts were stripped and clipped to 2000 characters before the prompt was prepended.
• Script: prepare_mix.py in the embeddingmagibu2 project.

Usage

from datasets import load_dataset

ds = load_dataset("alibayram/embeddingmagibu2-distill-mix", split="pct01")
row = ds[0]
row["a"], row["a_emb"][:4]

License

Each row keeps the license of its source dataset (linked above); check them before redistribution or commercial use. The vectors were produced with google/embeddinggemma-2 under its own terms of use.