Hugging Face's logo

Datasets:
KRLabsOrg
/
acl-verbatim-spans

Tasks:
Question Answering
Token Classification
Text Retrieval
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
acl-anthology
extractive-qa
evidence-selection
semantic-highlighting
silver-labels
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
acl-verbatim-spans / README.md
adaamko's picture
adaamko
Update README.md
2283a67 verified
preview code
|
Raw History Blame Contribute Delete
7.74 kB
metadata 
license: cc-by-4.0
task_categories:
  - question-answering
  - token-classification
  - text-retrieval
language:
  - en
pretty_name: ACL-Verbatim Span Dataset
tags:
  - acl-anthology
  - extractive-qa
  - evidence-selection
  - semantic-highlighting
  - silver-labels
configs:
  - config_name: canonical
    data_files:
      - split: train
        path: canonical/train-*
      - split: validation
        path: canonical/validation-*
      - split: test
        path: canonical/test-*
  - config_name: encoder
    data_files:
      - split: train
        path: encoder/train-*
      - split: validation
        path: encoder/validation-*

ACL-Verbatim Span Dataset

KRLabsOrg/acl-verbatim-spans is a dataset for query-conditioned extractive evidence selection over papers from the ACL Anthology.

The release combines:

  • a gold test benchmark with manual span annotations
  • a larger silver training set produced from synthetic questions, retrieval, and LLM-based span annotation
  • an encoder-ready config for training token-classification models directly

The underlying document collection is KRLabsOrg/acl-anthology-md.

What This Dataset Is For

This dataset is intended for systems that, given a question and a retrieved paper chunk, must identify the supporting evidence verbatim in the chunk.

Typical uses include:

  • training token classifiers for semantic highlighting
  • evaluating span extractors and evidence selectors
  • comparing LLM teachers, token-level students, and sentence-selection baselines
  • studying paragraph-scale evidence extraction in scientific text

Configs

canonical

One row per (question, chunk) pair.

  • train and validation are silver supervision
  • test is the manual gold benchmark

This is the main config for analysis, evaluation, and downstream reuse.

Current split sizes:

split rows notes
train 20,916 silver
validation 2,319 silver dev
test 100 gold benchmark

Expected fields:

field type notes
question string Query / question text
paper_id string ACL Anthology identifier
chunk_index int Chunk number within the paper
chunk string Raw chunk text
label int 1 if answer-bearing, 0 otherwise
answerable bool Teacher answerability decision
spans list[struct] {start, end, text} evidence spans
source string Where the candidate chunk came from in data generation: gold = the original chunk used to create the synthetic question, retrieved = a chunk retrieved for that question
retrieval_rank int or null Rank among retrieved candidates
gold_paper string Paper from which the synthetic question was derived
gold_chunk int Chunk from which the synthetic question was derived
predicted_texts list[string] Raw teacher outputs before alignment
latency_s float Teacher latency metadata
err string or null Teacher/extraction error, if any

encoder

Token-classification-ready rows derived from the silver canonical data by tokenization and windowing.

This config is intended for direct encoder training with Hugging Face transformers.

Current split sizes:

split rows
train 21,099
validation 2,343

Expected fields:

field type notes
input_ids list[int] Token ids
attention_mask list[int] Attention mask
labels list[int] Binary or BIO token labels

How the encoder split was built

The encoder split was pretokenized with:

parameter value
tokenizer answerdotai/ModernBERT-base
max_length 8192
doc_stride 256
truncation only_second (question kept whole, chunk windowed)
label_scheme binary (0 = outside, 1 = evidence)
drop_unlabeled_positives true

If you want to train with a different tokenizer or label scheme, rebuild from the canonical config:

python acl_verbatim/span_training/prepare_token_cls_dataset.py \
  --input-file <canonical_train.jsonl> \
  --output-file train.my_tokenizer.binary.jsonl \
  --tokenizer <your-tokenizer> \
  --label-scheme binary \
  --drop-unlabeled-positives

Evaluation always uses canonical/test (raw text); the encoder config intentionally does not include a test split, because span-level scoring needs to map token predictions back to character offsets in the original chunk, which requires retokenizing at inference time.

Annotation Convention

The benchmark uses paragraph-oriented evidence annotation rather than minimal SQuAD-style answer spans.

Important consequences:

  • broader supporting passages are often preferred over minimal snippets
  • tables, figure captions, and other structured evidence are considered valid positives
  • bibliography/reference sections are generally out of scope as positive evidence

Gold Benchmark Summary

The canonical/test split is the manually annotated benchmark used for extractor evaluation.

Queries 20
Retrieved chunks per query 5
Total rows 100
Relevant rows 47
Gold spans 78

Silver Training Data

The silver training data was produced by:

  1. sampling papers from the ACL Anthology corpus
  2. generating synthetic questions from paper chunks
  3. rewriting those questions into retrieval-style queries
  4. retrieving top-ranked chunks
  5. annotating answer-bearing chunks with an LLM span extractor
  6. filtering noisy positives while preserving table/caption evidence

Intended Uses

  • Train query-conditioned token classifiers for semantic highlighting
  • Compare LLM teachers, token models, sentence-level compressors, and semantic highlighters under one span-scoring harness
  • Study boundary conventions for evidence extraction in scientific text

Limitations

  • The gold benchmark is small and single-annotator
  • Silver labels inherit retrieval noise and teacher noise
  • Tables and captions are represented through markdown/caption text, not full table structure
  • The benchmark reflects a paragraph-oriented annotation convention rather than a strict minimal answer-span convention

How To Load The Dataset

Load the canonical config:

from datasets import load_dataset

ds = load_dataset("KRLabsOrg/acl-verbatim-spans", "canonical")
train = ds["train"]
dev = ds["validation"]
test = ds["test"]

Load encoder-ready training rows:

from datasets import load_dataset

encoder = load_dataset("KRLabsOrg/acl-verbatim-spans", "encoder")
train = encoder["train"]
dev = encoder["validation"]

Example: inspect gold vs silver splits 

from datasets import load_dataset

canonical = load_dataset("KRLabsOrg/acl-verbatim-spans", "canonical")

silver_train = canonical["train"]
silver_dev = canonical["validation"]
gold_test = canonical["test"]

Related Resources

Citation 

@misc{Recski:2026,
      title={ACL-Verbatim: hallucination-free question answering for research}, 
      author={Gábor Recski and Szilveszter Tóth and Nadia Verdha and István Boros and Ádám Kovács},
      year={2026},
      eprint={2605.21102},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.21102}, 
}