mjbommar/opengloss-tasks

OpenGloss Tasks (v2.1)

FLAN / GLUE / SuperGLUE-style instruction tasks built from our own seeds for training small language models (the ALEA models). Gold labels are deterministic wherever possible: document metadata, spans verified verbatim in the input, dictionary sense ids and relations, or construction. Where a task needs judgement (inference, similarity, paraphrase, reading comprehension, topic labels), labels come from two open-weight teacher models and are kept only when they agree; disagreements are dropped.

This version has 14 configs, 1,031,210 training documents and about 199.0M tokens (ALEA's 16K tokenizer).

CHANGES (v2 -> v2.1): tracking the stable.v4d mix

On 2026-10-12 (step 23,200) ALEA R3's stable phase switched to a new mix, stable.v4d (13.5B tokens; it trains until the stable phase ends at step 35,619). It removes a few layouts the model had over-learned, cuts the correction share, applies the latest evaluation references, and adds the NLP task suite. This version tracks it:

• The data is unchanged from v2: every file under data/ is byte-identical, and no row was removed.
• New file mapping/stable_v4d.parquet: one row per released document (and the few stable.v4d documents that are not released): id, text_sha1, config, stable_set, v4d_status, exposures (passes in stable.v4d), n_tokens_alea (as trained, incl. <|start|>/<|end|>), released (always true for released rows). v4d_status codes: trained (in the stable.v4d mix), kept_unplanned (in the set, but after the point where its planned fraction of a pass ends), removed_decon (selected for the mix, but left out of training because its text shares a word 8-gram with an ALEA evaluation item, mostly the private held-out task passages; 3,109 rows), removed_tox (toxicity score >= 0.8; 226 rows), eval_overlap_screen (not in the mix, and its text shares such an 8-gram; 2,402 rows) and not_in_v4d (not in the mix).
• The overlap codes only record that a row shares text with ALEA's evaluation sources (often a public-record passage that also appears there); the rows themselves are ordinary training data and stay published.

config	rows	trained in stable.v4d	in the set, unplanned	removed_decon	eval_overlap_screen	exposures per trained document in stable.v4d
acceptability	201,877	87,498	7,606	223	225	1.00
d11_register	32,438	0	0	0	6	-
d12_subject	88,214	0	0	0	21	-
d13_level_pair	40,577	0	0	0	8	-
d13_reading_level	34,528	0	0	0	10	-
gloss_paraphrase	111,682	65,329	5,844	631	294	1.00
lexical_relation	109,223	50,943	4,677	130	108	1.00
nli	30,726	0	0	0	33	-
sentiment	18,723	0	0	0	8	-
sts	17,839	0	0	0	458	-
synonym_substitution	56,848	27,704	2,478	16	11	1.00
wic	194,486	112,096	10,075	533	354	1.00
winograd	4,141	0	0	0	4	-
wsd	190,337	110,202	9,860	1,576	862	1.00

The NLP suite enters the stable mix for the first time in stable.v4d (1.5% of it): the deterministic configs' train renders, one render per item first; the teacher-labelled configs are not in it. Their rows are not_in_v4d. Rows of the decay split are never in the stable mix.

Configs

config	task	gold	rows	generator
acceptability	is the sentence well formed (grammar and spelling)?	construction: clean OpenGloss / code sentences vs alea-data-generator corruptions and grammar edits	170,471	deterministic (no model)
d11_register	register of a passage: plain, informal, formal or technical	the register each OpenGloss-Edu passage was written in	32,438	deterministic (no model)
d12_subject	subject area of OpenGloss example sentences or an OpenGloss-Edu passage	OpenGloss domain / OpenGloss-Edu subject metadata	88,214	deterministic (no model)
d13_level_pair	which of two passages on the same subject is easier (or harder) to read	the levels the two OpenGloss-Edu passages were written for	40,577	deterministic (no model)
d13_reading_level	reading difficulty of a passage: easy, medium or hard	the level each OpenGloss-Edu passage was written for	34,528	deterministic (no model)
gloss_paraphrase	are two definitions paraphrases of each other?	same OpenGloss sense (renditions) vs another sense of the same word	101,581	deterministic (no model)
lexical_relation	relation between two words (synonym, antonym, hypernym, hyponym)	OpenGloss typed relations, resolved, confidence >= 0.95	93,993	deterministic (no model)
nli	NLI on everyday premises	hypotheses written by one teacher, labels confirmed blind by another	30,726	hypotheses written by GLM-5.3-Flash; labels confirmed blind by DeepSeek-V4.1-Flash
sentiment	sentiment of an everyday sentence (positive / negative / neutral)	agreement of two teacher models	18,723	labels: GLM-5.3-Flash and DeepSeek-V4.1-Flash, agreement only
sts	semantic similarity of two sentences on a 0-5 scale	two teacher scores within 1 point; gold = their mean	17,839	scores: GLM-5.3-Flash and DeepSeek-V4.1-Flash, absolute difference <= 1, gold = mean
synonym_substitution	pick or give a synonym that fits a sentence	OpenGloss synonym relations of the example's own sense	48,596	deterministic (no model)
wic	word in context: same sense in two sentences?	OpenGloss sense ids of the example sentences	176,329	deterministic (no model)
winograd	Winograd-style pronoun resolution	schemas written by one teacher, both twin sentences resolved blind by two teachers	4,141	schemas written by GLM-5.3-Flash; each sentence resolved blind by DeepSeek-V4.1-Flash and GLM-5.3-Flash
wsd	choose or write the gloss of a word as used in a sentence	OpenGloss sense id of the example	173,054	deterministic (no model)

Format

Each row is one training document (text) plus structured fields. Documents come in two layouts:

• chat: ALEA chat tokens (<|start|>, <|system|>, <|user|>, <|assistant|>, <|eot|>, <|end|>); messages holds the same turns as a JSON list. The user turn always starts with the instruction (and any answer options), followed by the input; few-shot examples are earlier turns.
• plain: one instruction first, then each input in a delimiter block followed directly by its answer.

An instruction or label never follows the input. About half the documents are zero-shot and half few-shot; answer options are shown in about half of the documents of multiple-choice tasks, in shuffled order. Answers are bare labels, short sentences, or a short rationale inside <|think|> ... <|/think|> followed by the answer (reasoning-flavoured tasks only). The answer's form follows the instruction: an instruction that asks for working, steps or a reason always gets a rationale, and one that asks for only the label or number gets a bare answer. Instruction wording is drawn from about ten hand-written templates per task with combinatorial variations (plus, for the configs carried over from v1, template paraphrases written by GLM-5.3-Flash and kept only if every placeholder survived and no evaluation instruction is echoed); no opening, system prompt or answer framing exceeds 1% of a config's documents.

Several configs share source passages on purpose: the same passage appears under different instructions with different answers (for example the subject, reading level and register of one OpenGloss-Edu passage), so the instruction, not the passage, decides the answer.

Columns: id, task, subtask, format, shots, options_shown, answer_style, text, messages, label, provenance (seed source ids), generator. Configs carried over from v1 also have a decay split: the unchanged v1 documents of a smaller slice rendered with hand-written templates only, kept for reproducibility; train is re-rendered in v2.

Sources

Seeds are OpenGloss v2.4 (CC-BY-4.0): senses, glosses and gloss renditions, example sentences with target spans, and typed relations; and OpenGloss-Edu (CC-BY-4.0) passages with their level, register and subject metadata. Senses on a sensitive-sense list are excluded. The acceptability task also uses public-domain U.S. Code / eCFR sentences.

Teacher-labelled configs

Teacher models: GLM-5.3-Flash and DeepSeek-V4.1-Flash (open-weight), called through third-party hosts with zero data retention, never the model makers' own APIs. Gold rules:

• sentiment, sts, nli, winograd: labelled from OpenGloss example sentences only. Sentiment: both models agree; similarity: both scores within 1 point, gold = mean; NLI: hypotheses written by one model, labels confirmed blind by the other; Winograd: twin sentences written by one model, each resolved blind by both.

Decontamination and filters

Every document passed, fail-closed:

• n-gram decontamination against ALEA's evaluation bundle (benchmark items, held-out probes, template-leak prompts), every ALEA in-house evaluation set (in-context-learning suites, correction, step-scaling, spot-check, closed-book and chat probes), held-out word pairs and arithmetic operand pairs, and 13-gram hashes of public copyright-memorization evaluation sets; for public-record seed text, n-grams that occur in 10 or more seed documents count as boilerplate (statutory and regulatory formulae) and do not drop a document;
• an instruction-template guard: no task template contains an evaluation instruction;
• a toxicity filter (s-nlp/roberta_toxicity_classifier, any 64-word chunk >= 0.8 is dropped);
• render checks: no unfilled template placeholder, no repeated courtesy word or lead-in in one instruction, the answer's form matches what the instruction asks for, no stray control tokens, at most one Question/Answer label pair, no grade-level ladders.

Held-out evaluation splits (separate source documents, rendered only with separate held-out templates) exist for every config and are not released.

Changes

• v2: 4 new configs (subject area, reading level, reading-level comparison, register). All v1 configs re-rendered with the v2 render checks (answer form follows the instruction; no repeated courtesy words such as "Please do this item, please."; fixed subject-verb agreement in some templates).
• v1: first release.

License

The task data (templates, labels, renderings) is released under CC-BY-4.0. OpenGloss and OpenGloss-Edu are CC-BY-4.0.

Citation

Please cite the ALEA Institute and the OpenGloss / KL3M projects when using this data.