AbhishekBarali/padding-to-k

Padding to k

Benchmark data and scored results for the paper Padding to k: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists (Abhishek Barali, Aayush Paudel, 2026).

Code, preregistration and full record: github.com/AbhishekBarali/padding-to-k · Preregistration: osf.io/e4r9f

In short

Extraction pipelines ask a model for JSON that follows a schema, and the schema can set a minimum number of list entries ("minItems": 3). When the document holds fewer items than that minimum, most open-weight models pad the list to reach it, mostly with real text copied from the document. A minimum shown in the schema, but not enforced, raises the share of lists longer than the true count from 34.5% to 55.4%.

Every document here comes from a published, human-annotated dataset, so the true number of items (n) is known. Each one is asked for k = 1, 3 and 5 items under eight prompt conditions.

Load it

from datasets import load_dataset

prompts   = load_dataset("AbhishekBarali/padding-to-k", "prompts", split="test")    # 24,144 prompts
documents = load_dataset("AbhishekBarali/padding-to-k", "documents", split="test")  # 503 documents
results   = load_dataset("AbhishekBarali/padding-to-k", "results", split="test")    # 494,551 scored responses

prompt is the exact user message the models received, so you can send it to your own model as it is.

Configs

prompts (default)

One row per test prompt.

column	meaning
request_id, doc_id	prompt id and the document it uses
split	pilot (60 documents) or full
set	which document set (see below)
condition	prompt condition (see below)
k, n	items asked for, and true items in the document
wording_id, instruction	the instruction wording, copied from real-world prompts
json_schema	the JSON schema, as a JSON string
enforce_schema	true if the decoder should enforce the schema (constrained decoding)
prompt	the full rendered user message
ideal_output	the correct answer, as a JSON string
any_k_of_gold, placeholders_ok, empty_ok, padding_slots	what else counts as correct (e.g. placeholders for missing items)

Conditions:

condition	what the model gets
none	"List all the ..." and a schema with no count
prose_up_to_k	"Extract up to k ..." (5 real wordings)
prose_exactly_k	"Extract exactly k ...", "Return the top k ..." (5 real wordings)
max_k_shown	schema with maxItems: k, shown in the prompt
min_k_shown	schema with minItems: k, shown but not enforced
min_k_shown_fewer_ok	min_k_shown plus "If the document contains fewer than k, list all available ones."
min_k_enforced	same schema, enforced by the decoder
min_k_escape	minItems: k enforced, plus "If the information is not available, use N/A."

documents

One row per study document: the text, the true items (gold_items, with character offsets where the item is an exact quote; meeting action items are summary sentences and carry the transcript lines they point to in evidence), the source dataset and a pointer back into it (source, source_id), the licence, and flags recording any change to the text (for example a removed "TOTAL ITEMS 3" line on a receipt).

set	one document is	list asked for	source
meetings	a window of a meeting transcript	action items	AMI Meeting Corpus
emails	one Enron email	requests	MailEx
receipts	one receipt as plain text	items purchased	CORD v2
job_posts	sentences from one job ad	skills	SkillSpan
job_posts_knowledge	sentences from one job ad	tools, technologies, fields	SkillSpan
clinical	sentences from a published case report	medications	MACCROBAT2020
wiki_people	one Wikipedia sentence	people	Few-NERD
novel_quotes	1-3 paragraphs of a novel	direct speech	PDNC

The study also uses 55 product reviews from ROAST. Their text has no stated licence, so it is not included here; data/build.py in the GitHub repo rebuilds them from the original source.

results

One row per scored model response from the study runs: eight open-weight models served on a GPU and six more on hosted services (hosted = true). run_kind is full (the main grid), repeat (a repeated subset), reask (re-asking with the validation error), think (thinking on/off comparisons) or dev (matcher development on reserve documents, not in the study). Columns include whether the output was valid JSON, the number of entries and substantive entries, over_count (more substantive entries than true items), exact_k, and the matcher's per-entry counts (matched_gold, n_wrong_type, n_invented, ...). Per-entry labels are in the GitHub repo (analysis/full/scored/). How each number is computed is in the paper.

Licence

CC BY-SA 4.0, because three source datasets are share-alike. Please also cite the source datasets (listed with their licences in LICENSE-DATA.md). The Enron emails are real people's mail, public since 2003; the case reports are published and de-identified.

Citation

@misc{barali2026padding,
  title         = {Padding to $k$: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists},
  author        = {Barali, Abhishek and Paudel, Aayush},
  year          = {2026},
  eprint        = {2610.30997},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.30997}
}