jungseob/directional-verification

Directional Verification

Data for Distilling Directional Verification. The code is at github.com/js-lee-AI/directional-verification.

A teacher may recall a relation in one direction and still fail to generate the answer in the reverse direction. The same teacher can often recognize the right answer when it scores the relation in the direction it knows. The paper turns that check into training labels for parent to child questions and trains masked diffusion students on them.

This dataset holds what those experiments read and produce: the Wikidata parent and child facts, the candidate pools of three cohorts, the scores of four frozen teachers in both directions, and the answers of the trained students.

Subsets

subset	splits	rows	one row is
pools	screened, uniform64, lexical64	1,500 / 512 / 512	a parent to child query with its candidates and its pseudo-label
scores	screened, uniform64, lexical64	6,000 / 2,048 / 2,048	the scores of one teacher for the candidates of one query
predictions	screened, uniform64, lexical64	16,680 / 3,072 / 3,072	the answer of one trained student to one query
corpus	train	8,218	a child with its recorded parents

The splits are the three cohorts of the paper. screened is the screened cohort with its acquired candidate pools. uniform64 and lexical64 are the unscreened cohort, the same 512 queries with 64 uniformly sampled names or 64 lexically similar names per query.

Quick start

from datasets import load_dataset

pools = load_dataset("jungseob/directional-verification", "pools", split="screened")
print(pools[0]["parent"], "->", pools[0]["label"])
# Munir Ertegun -> Nasuhi Ertegun

The known-direction labels can be rebuilt from the scores with numpy alone.

import numpy as np
from datasets import load_dataset

repo = "jungseob/directional-verification"
scores = load_dataset(repo, "scores", split="screened")
corpus = load_dataset(repo, "corpus", split="train")

children = {}
for row in corpus:
    for parent in row["parents"]:
        children.setdefault(parent, set()).add(row["child"])

by_query = {}
for row in scores:
    by_query.setdefault(row["id"], []).append(row)

correct = 0
for rows in by_query.values():
    parent, candidates = rows[0]["parent"], rows[0]["candidates"]
    mean = np.mean([row["known_mean"] for row in rows], axis=0)  # four teachers, eq. 2
    mean[[name == parent for name in candidates]] = -np.inf      # never the queried parent
    correct += candidates[int(mean.argmax())] in children[parent]
print(f"{100 * correct / len(by_query):.2f}")
# 90.73

This is the 90.73 of the Known column in Table 4 of the paper.

Columns

pools

column	type	meaning
id	string	query ID, such as screened_0000
parent	string	the queried parent
child	string	the recorded child, used only for evaluation
candidates	list of strings	candidate children, in the order the scores follow
label	string	the known-direction pseudo-label, the candidate with the highest four-teacher mean score other than the queried parent
evaluated	bool	true for the exposure-free queries the students are scored on, 1,390 in screened and all 512 in the other two

scores

One row per query and teacher, four rows per query. Every list follows the order of candidates. Scores are log probabilities.

column	type	meaning
id, parent, candidates		as in pools
teacher	string	qwen, olmo, mistral or llama
known_mean	list of floats	mean-token score of the parent after {child}'s parent is, for each candidate child (eq. 1)
reverse_mean	list of floats	mean-token score of each candidate after {parent}'s child is
reverse_tokens	list of ints	token count of each candidate in the reverse prompt
parent_tokens	int	token count of the parent in the known-direction prompt
context_mean	list of floats	mean-token score of each candidate after the context-only prompt The child is
context_tokens	list of ints	token count of each candidate in the context-only prompt

A summed-token score is the mean-token score times its token count. The DC reverse score of eq. 3 is the reverse score minus the context-only score, in the same form for both.

teacher	model	revision
qwen	Qwen/Qwen3-8B	b968826d9c46dd6066d109eabc6255188de91218
olmo	allenai/OLMo-2-1124-7B	7df9a82518afdecae4e8c026b27adccc8c1f0032
mistral	mistralai/Mistral-7B-v0.3	caa1feb0e54d415e2df31207e5f4e273e33509b1
llama	meta-llama/Llama-3.1-8B-Instruct	0e9e39f249a16976918f6564b8830bc894c89659

predictions

column	type	meaning
id	string	query ID, one of the evaluated queries of the split
label_source	string	the labels the student was trained on, known_mean, dc_sum, dc_mean or reverse_mean
seed	int	0, 1 or 2, which the paper calls runs A, B and C
text	string	the student's answer

The students are masked diffusion models that start from Qwen/Qwen3-0.6B. These are the runs behind Table 1 of the paper.

corpus

column	type	meaning
child	string	a child
parents	list of strings	the recorded parents of the child
withheld_screened, withheld_uniform64, withheld_lexical64	bool	true when the forward facts of this child are removed before the students of that cohort are trained

The corpus has 10,505 parent and child pairs over 8,218 children and 9,918 parents.

Raw files

raw/ holds the files of the data/ folder of the code repository, unchanged. The table scripts and the ddv library read this layout.

hf download jungseob/directional-verification --repo-type dataset --include "raw/*" --local-dir ddv-data

import ddv

pool = ddv.load_pool("screened", "ddv-data/raw")
channels = ddv.load_scores("screened", "ddv-data/raw")["known_mean"]

Source

The facts were mined from the public Wikidata Query Service on July 12, 2026. Each query selects humans with an English Wikipedia article and one of 19 occupations, together with their recorded fathers or mothers, and returns English labels. Names are accent-stripped and kept when they consist of two to four ASCII tokens. People are identified by these cleaned names.

Considerations

Every name is a public Wikidata entry, and every pair in corpus is a relation recorded in Wikidata. The label column of pools and the text column of predictions are model outputs. Many of them name a person who is not a child of the queried parent, and a few student answers are garbled names. Do not read them as facts about the people they name.

License

The files are released under the MIT license, as in the code repository. The Wikidata facts are CC0. The scores are outputs of the four teachers, so check each teacher's license before you distribute a model trained with them. The Llama 3.1 Community License in particular sets conditions on such models.

Citation

@article{lee2026distilling,
  title   = {Distilling Directional Verification},
  author  = {Lee, Jungseob and Eo, Sugyeong and Hong, Seongtae and Lee, Seungyoon and Park, Chanjun and Seo, Jaehyung and Lim, Heuiseok},
  journal = {arXiv preprint arXiv:2610.00997},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.00997}
}