caiovicentino1/ekbasis-data

Ekbasis Data

Papers: Look When Unsure, Check When Sure · DOI 10.5281/zenodo.23146970 · When Does a Consequence Model Make AI Agents Safer? · DOI 10.5281/zenodo.23197341 · Code: github.com/OpenInterpretability/ekbasis · Site: openinterp.org/ekbasis

The data behind Ekbasis-27B, an open world model for agents: the rows its whole training lineage was trained on, and the sets it was evaluated on. Every answer in the generated rows was computed, not written: by a simulator for the worlds, by running real git in a throwaway repository for git.

What is inside

Config	Splits (rows)	What
lineage	train 198,721	Every row the lineage's training files held (the V42 file holds all earlier ones; r4a adds two mined sets)
guard	known 4,486, held 6,394	The larger guard set: 'Does it lose uncommitted work?' and 'Does it fail?' on fresh git scenarios (real git), known command types and held-out ones
git3	known 1,482, held 1,623	The git test sets of the release evaluation (real git), known command types and held-out ones
worlds	multi_testfam 600, multi_trainfam 600, ftest_family 1,080	Simulated worlds: several questions on one state (families trained and never trained) and never-trained families
confident_errors	test 15,008	The pre-registered confident-errors test: 15,008 questions in 12 world families
git2	known 1,545, held 1,622	Git v2 test sets (the 240-question comparison is drawn from them)
validation	m3val 300, fval_git 1,390, fval_git2 1,514, fval_git3 1,524	The validation sets the lineage chose its checkpoints by

Every row has: id (a hash of the row), kind (the training pool it was drawn from), stages (training rows: the lineage stages whose training file held it), world, qtype, steps (actions in the prompt), prompt (the situation as the model reads it: the rules, the state, the actions), question (the typed question as JSON: yes/no, or a choice with its options), answer, license, source (the generator, or the upstream row), and item (the original row as JSON, with the hindsight prompt ceiling used by the training objective and the answer before the actions, gold0). raw/ holds the exact JSONL files the training and evaluation code read.

from datasets import load_dataset
train = load_dataset("caiovicentino1/ekbasis-data", "lineage", split="train")
guard = load_dataset("caiovicentino1/ekbasis-data", "guard")   # splits: known, held

The training lineage

The lineage has six training stages, each continuing the last. The first five each hold the previous stage's file. The sixth, r4a, continued V42 on V42's file plus two sets mined from V42's own answers. lineage/train holds every row any stage saw, and stages says which stages held each row.

Stage	From	Training file	Mix (probability of each kind per batch item)	Steps × batch	LR	Checkpoint chosen by
27s1	Eikos-27B	ftrain_m3only (24,003 rows)	floor (factory 3, labeled floor3 later) 0.85, read 0.15	1000 × 8	5e-05	m3val
V35	27s1 (step 1000)	ftrain_mixG (118,218 rows)	floor 0.15, floor3 0.15, floorG 0.3, read 0.05, replay 0.35	600 × 8	1.5e-05	fval_git
V37	V35 (step 600)	ftrain_mixG2 (149,591 rows)	floorG2 0.4, floorG 0.1, floor 0.1, floor3 0.1, read 0.05, replay 0.25	600 × 8	1.5e-05	fval_git2
V41	V37 (step 600)	ftrain_mixG2M (173,591 rows)	multi 0.4, floorG2 0.15, floorG 0.05, floor 0.1, floor3 0.08, read 0.04, replay 0.18	400 × 8	1.5e-05	fval_git2
V42	V41 (step 400)	ftrain_mixG3 (198,276 rows)	floorG3 0.35, floorG2 0.1, floorG 0.05, multi 0.15, floor 0.1, floor3 0.07, read 0.04, replay 0.14	400 × 8	1.5e-05	fval_git3
r4a	V42 (step 400)	ftrain_mixG3 + mined_wide + mined_chain (206,289 rows)	floorG3 0.245, floorG2 0.07, floorG 0.035, multi 0.105, floor 0.07, floor3 0.049, read 0.028, replay 0.098, floorM 0.2, floorC 0.1	300 × 8	1e-05	fval_git3

LoRA rank 64 (alpha 128) on every linear layer; cross-entropy on the letter readout plus a JEPA-style term (weight 3, layers 16, 32, 48, 64, at the end of the evidence); each batch item drawn from a kind with the mix's probability. Full recipe: recipes.json.

The release is not a training stage. Ekbasis-27B is the exact interpolation halfway between the V42 and r4a adapters: the two LoRA adapters concatenated into one of rank 128, which gives the weights W0 + ½ ΔW(V42) + ½ ΔW(r4a); the result was merged. Why and how it was chosen: RELEASE_EVAL.md and the paper Look When Unsure (paper/) in the repository.

Kinds: floor (simulated worlds, factory 1), floor3 (factory 3), read (reading a true outcome: the hindsight prompt as input), multi (several questions on one state, read once), floorG, floorG2, floorG3 (git v1, v2, v3: real execution as an unprivileged user in throwaway repositories with synthetic contents and authors), replay (items of the Eikos training data, mixed in so the base model's decision skills do not narrow), floorM (new world items that V42 answered wrong or right but below 0.99, from true states) and floorC (items mined inside V42's own chains: each variable it got wrong or answered below 0.99, with the true result of the action from the state it was reading). In the mined sets the model only chose which items enter; every answer is the simulator's.

What is not here, and why

• Replay rows not redistributed: 7,423 (7,348 distinct). The replay rows come from the Eikos training data; only those that are rows of the published eikos-decisions are redistributed here, each with its eikos-decisions id and upstream license in source and license. The others were left out of that release by its per-source quotas, or excluded by it (443, among them 151 judging items built on FinQA, a dataset Eikos keeps for evaluation; none of Ekbasis' evaluations uses it). They are listed by id in checks.json.
• Rows flagged by the privacy scan: 19. Every row was scanned with the Eikos release's scanner (credentials, e-mail addresses at consumer domains, checksum-valid payment cards, SSNs, IBANs, CPFs); a flagged row is withheld, never edited, and listed by id in checks.json. All of them were replay rows that are not redistributed anyway.

Checks

• Overlap between training and evaluation (corrected 2026-10-06). No evaluation or validation row is also a training row as a whole (the same prompt, question and answer), but the world generator's small worlds repeat across seeds, so many world prompts repeat or nearly repeat a training prompt. Checked against every lineage/train row (an identical prompt, or a near duplicate: word 8-gram Jaccard ≥ 0.5 against one row, after removing each set's own layout):

  • confident_errors: 5,378 of 15,008 items (174 identical prompts, 5,204 near duplicates; 4,151 of 7,172 in the trained families, 1,227 of 2,472 with held-out question types, none in the never-trained families). 104 share the prompt and the question with a training row (all in the sequences family, all answered right by the model); the first version of this card counted 88;
  • worlds/multi_trainfam: 291 of 600 states (5 identical prompts, with the same questions);
  • git3, git2, guard, worlds/multi_testfam, worlds/ftest_family and validation: none (4 fval_git items have a hindsight prompt equal to a training row's).

  Every published number, recomputed without the overlapping items, is in results/overlap_audit/ of the code repository: every pre-registered conclusion holds. The guard sets were generated after removing every state that appears in any training, validation or test file, from seeds no git generator had used.

• No outputs of Claude, Jev or any other closed model are in the training rows. The replay items were generated by programs, written by Qwen3.8-27B or GLM-5.3-Flash, or built from public datasets, and labeled by GLM-5.3-Flash (MIT, Z.ai); see the eikos-decisions card.

Licenses and attribution

The generated rows (worlds and git) are released under Apache-2.0. The replay rows keep their own license, given in the license column of each row: CC BY 4.0 (eikos-decisions), with these upstreams:

Upstream	License	Credit
FinEntity	ODC-BY 1.0	Tang et al. (2023), FinEntity: Entity-level Sentiment Classification for Financial Texts, EMNLP 2023
TAT-QA	CC BY 4.0	Zhu et al. (2021), TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance, ACL 2021
GSM8K (train)	MIT	Cobbe et al. (2021), Training Verifiers to Solve Math Word Problems
eikos-decisions	CC BY 4.0	Vicentino (2026), Eikos Decisions

Contact

Questions about the data, or a row that should not be here: caio@openinterp.org.

Citation

@misc{ekbasis2026data,
  title  = {Ekbasis Data: the training lineage and evaluation sets of an open consequence model},
  author = {Caio Vicentino},
  year   = {2026},
  url    = {https://huggingface.co/datasets/caiovicentino1/ekbasis-data}
}