theaviv/lev1-decisions

Lev1 decisions

What is this?

Synthetic decision data from Lev1-3.7B, a small model that scores every option of a typed question in one forward pass and returns probabilities, not text. A record has a state (a document, a table, a policy, a list of tools), a set of typed questions about it, and the gold answer to each. A question is either a choice among options or a yes/no check. There are 20,053 records in 21 configs.

How it was made

17 configs come from code generators. Each one draws a random scenario, computes the answer with ordinary code, and writes near-miss distractors. The generators are not part of this release; the data is.

4 configs are text written by a language model. MiMo-V2.6-Flash wrote each text from a recipe: a label plus a few fixed signals, kept in meta.recipe. Gemma-4-31B-it and Qwen3-235B-A22B-2507 then labelled the text without seeing the recipe, and a text was kept only where the checkers agreed with its label. meta.p_gold holds the probability each checker that scored it gave the gold label.

Splits follow the training run. validation is drawn from the same templates as train. test uses templates that train never sees, and only some generator configs have one.

Before release I checked every record against the private test set of the Decision Index, with exact string matching, a 13-word overlap rule and a scan for its long documents. None of the 20,053 records matched.

Configs

Config	What it tests	train	validation	test	Licence
gen_arithmetic	Order totals, a budget check, the costliest line	386	14	120	CC-BY-4.0
gen_boolean	Nested and/or/not over true and false	712	37	30	CC-BY-4.0
gen_catalog	Cheapest or priciest item among 100 to 255 catalog options	613	27	40	CC-BY-4.0
gen_causal	Association versus intervention in confounded rates	939	61	40	CC-BY-4.0
gen_code	Output of a short Python function	941	59	40	CC-BY-4.0
gen_compare	Pairwise and maximum comparisons over people	334	16	60	CC-BY-4.0
gen_dates	Day counts, a return window, the weekday	370	24	120	CC-BY-4.0
gen_deduce	Ordering puzzles from true clues	950	50	40	CC-BY-4.0
gen_facts	Two-hop lookup: person, employer, head-office city	852	44	59	CC-BY-4.0
gen_facts_long	The same lookup hidden in long distractor text	239	11	-	CC-BY-4.0
gen_home	Which rule of a home hub fires first, and whether another holds	946	54	40	CC-BY-4.0
gen_liars	Truth-teller chains	749	47	38	CC-BY-4.0
gen_navigate	Step-and-turn routes: do you end back at the start	752	42	32	CC-BY-4.0
gen_policy	Return-policy rules with exceptions and missing facts	1,200	51	113	CC-BY-4.0
gen_table	Counting and ranking rows of a small table	748	52	40	CC-BY-4.0
gen_tools	Choosing the next tool from a synthetic API catalog	895	55	44	CC-BY-4.0
gen_track	Who holds what after a series of swaps	942	58	40	CC-BY-4.0
pack_halluc	Is a response supported by a Wikipedia passage	1,329	61	-	CC-BY-SA-3.0
pack_phish	Phishing or genuine email, and does the sender match the brand	1,426	69	-	CC-BY-4.0
pack_sarcasm	Sarcastic or sincere short post	1,493	87	-	CC-BY-4.0
pack_stance	Stance of a post on a local topic	1,355	67	-	CC-BY-4.0

Format

Each config is a set of zstd-compressed Parquet files, one row per record. Every column is a string. The state, questions, gold and meta columns hold JSON, so decode them after loading:

import json
from datasets import load_dataset

ds = load_dataset("theaviv/lev1-decisions", "gen_policy", split="train")
row = ds[0]
state, questions, gold = (json.loads(row[k]) for k in ("state", "questions", "gold"))

A state may be a string, an object or a list. Option keys can be opaque (k_6836); the description carries the meaning. One row, with long values cut:

{
  "id": "gen_policy:0",
  "source": "gen_policy",
  "family": "synthetic",
  "licence": "CC-BY-4.0",
  "state": "\"Policy: Rules: (1) final sale means no return; (2) returns close 30 days after delivery; (3) opened goods need proof of purcha...",
  "questions": "{\"eligible\": {\"type\": \"choice\", \"instructions\": \"Can this item be returned under the policy?\", \"criteria\": {\"yes\": \"It can be r...",
  "gold": "{\"eligible\": \"no\"}"
}

meta is present on the pack_* configs only. gold maps each question name to its right option key, or to yes or no.

Licence and attribution

The generator configs and pack_phish, pack_stance and pack_sarcasm are CC-BY-4.0. Credit "Lev1 decisions by Aviv Dozorets". The LLM-written text came from MiMo-V2.6-Flash (MIT licence), and the label checkers are Apache-2.0 models.

pack_halluc is CC-BY-SA-3.0. Its passages are Wikipedia text, so it needs attribution and share-alike. Each record names its article and URL in meta.

Related: Lev1-3.7B is trained on this data together with public datasets. Open-Jev is a CC0 dataset by another author in the same typed-decision format; it is not part of this release.

Limits

• It is templated. A model that learns these templates has not learned the world, so train on it alongside real data.
• The pack labels come from two models agreeing, not from people. Some are wrong, and the checkers' shared blind spots are in them.
• English only.
• Questions that had only a soft label were removed (unanswerable lookups in gen_facts and gen_facts_long), so every remaining question has a gold answer.