ogma-permissive-blend-v2

10.6M-text permissive-license blend for training small English embedding models. 14 sources, MTEB-contamination flagged, ready for teacher-embedding caching.

Motivation

v1 of the axiotic ogma small-embedding line was distilled from the CC-BY-NC jinaai/jina-embeddings-* teacher — great quality, but the non-commercial license contaminates any downstream commercial use. This blend drops all non-commercial data. Every source is under a permissive or share-alike licence (MIT, Apache-2.0, ODC-BY, CC-BY-4.0, CC-BY-SA-4.0). The industry convention followed here is that model weights trained on text are not themselves a derivative work of that text; we still surface the per-sample license so downstream users can make their own call.

The blend targets three things:

1. Broad domain coverage for a small (≤100M-param) English embedding model — web, wiki, QA, NLI, patents, news, forums, edu.
2. Semantic-similarity supervision via the QA / NLI / duplicate-question / STS sources.
3. Cleanly-flagged MTEB contamination so downstream MTEB numbers can be honestly reported.

Blend composition

#	Source	HF id	Samples kept	License
1	MS MARCO	microsoft/ms_marco	1,997,941	MIT
2	SNLI	stanfordnlp/snli	499,931	CC-BY-SA-4.0
3	MultiNLI	nyu-mll/multi_nli	499,999	OANC / CC-BY
4	GooAQ	sentence-transformers/gooaq	1,500,000	Apache-2.0
5	AG News	fancyzhx/ag_news	119,972	Academic (news headlines)
6	SQuAD v2	rajpurkar/squad_v2	149,206	CC-BY-SA-4.0
7	Wikipedia (en, 2023-11-01)	wikimedia/wikipedia	1,999,904	CC-BY-SA-4.0
8	C4 (en)	allenai/c4	1,500,000	ODC-BY
9	StackExchange (title-title pair)	sentence-transformers/stackexchange-duplicates	453,356	Apache-2.0 (repackaged)
10	STS-B	sentence-transformers/stsb	9,582	CC-BY-SA-4.0
11	FineWeb-Edu (sample-10BT)	HuggingFaceFW/fineweb-edu	999,920	ODC-BY
12	BigPatent (a subset)	big_patent	500,000	CC-BY-4.0
13	HotpotQA	sentence-transformers/hotpotqa	200,000	CC-BY-SA-4.0
14	Natural Questions	sentence-transformers/natural-questions	175,179	CC-BY-SA-4.0
Total			10,604,990

Exact per-source counts also live in combine_stats.json.

Schema

Single parquet file at data/blend.parquet, zstd-compressed, row groups of 50k:

column	type	notes
id	int64	Row id, dense [0, 10604989].
text	string	The text sample as fed to the tokenizer.
source	string	One of the 14 source keys in the table above.
role_hint	string	Coarse role tag from the ingest pipeline (text, query, doc, sym, etc.) — hints how the row is meant to be used but is not a hard label.
hash	string	BLAKE2b hex hash of the normalised text (used for dedup).
text_len	int32	Character length.
contamination_hit	bool	True iff this row's normalised hash matches a sample in the test set of one of the SMALL_MTEB_TASKS (see below).

Preprocessing

• Per-source sampling to the target counts above with seed=20260715.
• BLAKE2b hash dedup on lightly-normalised text (whitespace-collapsed, lower-cased).
• Length filter: 10 <= len(text) <= 8000 characters. No aggressive cleaning otherwise — this is meant to be a raw corpus, not a curated eval set.
• Dedup is applied both per-source and again on the combined pool, giving 10,604,990 unique rows out of 32.95M raw samples pulled.

MTEB contamination check

11,244 rows (0.11% of the corpus) match — by normalised-hash equality — a text that appears in the test set of one of the following 20 MTEB tasks:

AmazonCounterfactualClassification, Banking77Classification,
ToxicConversationsClassification, BiorxivClusteringS2S.v2,
MedrxivClusteringS2S.v2, TwentyNewsgroupsClustering.v2,
SprintDuplicateQuestions, TwitterSemEval2015, AskUbuntuDupQuestions,
SciDocsRR, NFCorpus, SciFact, ArguAna, STS12, STS13, STS15,
STSBenchmark, BIOSSES, SummEval, SummEvalSummarization.v2

These are flagged via the contamination_hit column so you can filter them out for any downstream MTEB-reported number:

from datasets import load_dataset
ds = load_dataset("axiotic/ogma-permissive-blend-v2", split="train")
clean = ds.filter(lambda r: not r["contamination_hit"])  # ~10,593,746 rows

The raw hit-id list is also shipped as contamination_hits.json.

Recommended usage

This corpus was built for embedding-model distillation — encode every text with a strong open teacher, then train a small student against the teacher's vectors (Matryoshka + InfoNCE + soft-cosine). A precomputed teacher-embedding cache using dunzhang/stella_en_1.5B_v5 in three modes (qry, doc, sym), fp16, L2-normalised, (10604990, 1024) per mode (~21 GB × 3 = ~63 GB), is available on request — email iam@antreas.io.

Suggested pipeline:

from datasets import load_dataset
ds = load_dataset("axiotic/ogma-permissive-blend-v2", split="train", streaming=True)
for row in ds:
    ...  # tokenise, encode with teacher, train student

License

• Aggregate release: CC-BY-4.0.
• Individual samples retain their upstream licences — see the "Blend composition" table. If you redistribute in a form other than model weights, honour those per-sample licences (e.g. share-alike attribution for CC-BY-SA rows).
• No PII scrubbing beyond what the upstream sources already do.

Citation

@misc{axiotic2026ogmapermissive,
  title        = {ogma-permissive-blend-v2: a 10.6M-text permissive-license blend for small English embedding models},
  author       = {Antoniou, Antreas and {Axiotic AI}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/axiotic/ogma-permissive-blend-v2}},
  note         = {Dataset release, Axiotic AI}
}