The Verbatim-RAG Extractor is a query-conditioned token classifier that highlights the verbatim spans of a passage that answer a question. It is the encoder companion to VerbatimRAG and the successor to verbatim-rag-modern-bert-v1. Built on Alibaba-NLP/gte-reranker-modernbert-base, which provides the long ModernBERT context (up to 8192 tokens) and a query-conditioned reranking prior on top of which span extraction is fine-tuned.

The goal is a lightweight extractor that can replace many LLM-based evidence highlighting calls in production RAG systems: local, deterministic, cheap to serve, and still competitive on span-overlap quality. In our ACL-Verbatim gold benchmark, the ACL-specialized sibling model is on par with strong LLM extractors by word-level F1, while this generic multi-domain model beats public extractive baselines across ACL gold, RAGBench, Squeez, and QASPER.

You can use it as the extraction stage inside VerbatimRAG, or drop it into your own RAG pipeline after retrieval/reranking to turn retrieved chunks into grounded evidence spans before displaying them to users or passing them to a generator.

Most public evidence extractors (Provence, Zilliz Semantic-Highlight, MultiSpanQA-trained models) are trained on Wikipedia-style prose QA only. This model is trained on KRLabsOrg/verbatim-spans, which adds financial tables, legal contracts, medical literature, product manuals, and — uniquely among public extractors — coding-agent tool output (pytest failures, git diff hunks, stack traces). The result is a single 150M-parameter encoder usable across the content shapes a real RAG or agent pipeline tends to retrieve, not just article paragraphs.

For an ACL-Anthology-specialized variant, see KRLabsOrg/acl-verbatim-modernbert.

Model Details

Architecture: ModernBERT (gte-reranker-modernbert-base) with 8192-token context
Task: Token classification — binary evidence labels mapped to character spans
Training Dataset: KRLabsOrg/verbatim-spans (multi-domain)
Language: English
Parameters: 150M

Training data composition

content shape	source
scientific paragraphs with citations	ACL silver
Wikipedia / general QA, multi-hop	RAGBench (HotpotQA, MS MARCO, ExpertQA, ...)
financial tables	RAGBench (TAT-QA, FinQA)
medical literature	RAGBench (PubMedQA, CovidQA)
legal contracts	RAGBench (CUAD)
product manuals	RAGBench (eManual, TechQA)
code, tool output, stack traces, logs	Squeez (SWE-bench tool outputs)

How It Works

A (question, context) pair is encoded as a single sequence; the model predicts a per-token positive-class probability over the context tokens. Above a threshold, contiguous positive runs are merged into character spans, with post-processing (min_span_chars, merge_gap_chars) that removes fragmentation artifacts. Long contexts are handled with sliding windows of max_length tokens stepped by doc_stride, and spans are merged across windows.

Usage

python

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "KRLabsOrg/verbatim-rag-modern-bert-v2",
    trust_remote_code=True,
)

result = model.process(
    question="What is ModernBERT?",
    context=(
        "ModernBERT is a long-context encoder for NLP. "
        "It supports sequences up to 8192 tokens. "
        "Unlike earlier BERT variants, it uses rotary position embeddings."
    ),
    threshold=0.2,
)

for span in result["spans"]:
    print(f"[{span['score']:.2f}] {span['text']}")

Use inside VerbatimRAG

python

from verbatim_rag.core import VerbatimRAG
from verbatim_rag.index import VerbatimIndex
from verbatim_rag.extractors import ModelSpanExtractor
from verbatim_rag.vector_stores import LocalMilvusStore
from verbatim_rag.embedding_providers import SpladeProvider

# v2 is the default ModelSpanExtractor model, but passing it explicitly makes
# the dependency clear.
extractor = ModelSpanExtractor(
    model_path="KRLabsOrg/verbatim-rag-modern-bert-v2",
    threshold=0.2,
    min_span_chars=30,
    merge_gap_chars=20,
    device=None,  # auto-detects cuda, then mps, then cpu
)

sparse_provider = SpladeProvider(
    model_name="opensearch-project/opensearch-neural-sparse-encoding-doc-v2-distill",
    device="cuda",  # use "cpu" if no GPU is available
)

vector_store = LocalMilvusStore(
    db_path="./index.db",
    collection_name="verbatim_rag",
    enable_dense=False,
    enable_sparse=True,
)

# Assumes the index has already been populated with your documents.
index = VerbatimIndex(
    vector_store=vector_store,
    sparse_provider=sparse_provider,
)

rag = VerbatimRAG(
    index=index,
    extractor=extractor,
    k=5,
)

response = rag.query("Main findings of the paper?")
print(response.answer)

You can also use the model directly after your own retriever/reranker:

python

from transformers import AutoModel

extractor = AutoModel.from_pretrained(
    "KRLabsOrg/verbatim-rag-modern-bert-v2",
    trust_remote_code=True,
)

question = "What evidence supports using DINOv2 as the visual backbone?"
context = (
    "We investigate different visual backbones for feature extraction. "
    "The results demonstrate DINOv2's effectiveness as a feature extractor "
    "for sign language translation."
)

result = extractor.process(question=question, context=context)

for span in result["spans"]:
    print(
        {
            "start": span["start"],
            "end": span["end"],
            "text": span["text"],
            "score": span["score"],
        }
        )

.process() accepts: question, context, threshold (default 0.2), max_length (default 8192), doc_stride (default 256), min_span_chars (default 30), merge_gap_chars (default 20), return_sentence_metrics (default False). For short-answer benchmarks (file paths, table cells, numbers), threshold=0.1 and min_span_chars=10 is the recall-tuned config documented in Performance below.

The return shape is {"spans": [{"start": int, "end": int, "text": str, "score": float}, ...]}, with "sentences" added when return_sentence_metrics=True. Spans are character offsets into the input context and are merged across sliding windows.

Performance

Evaluated with the shared span-extraction harness used by acl-verbatim. The current metric protocol scores every row in a slice: rows without gold spans are negative examples, and false-positive extracted text lowers precision.

The table below compares this generic model to two public extractive baselines: Zilliz Semantic Highlight and Provence. All systems are evaluated with the same all-row scorer. Latency is intentionally omitted because runtime depends on device, batching, and serving setup.

dataset	system	Word-P	Word-R	Word-F1	IoU@0.5	AnyOverlap	OverPred
ACL gold	verbatim-rag-modern-bert-v2	0.625	0.368	0.463	0.366	0.449	0.679
ACL gold	Zilliz semantic-highlight	0.470	0.221	0.301	0.113	0.513	1.500
ACL gold	Provence	0.276	0.457	0.344	0.153	0.718	3.013
RAGBench	verbatim-rag-modern-bert-v2	0.516	0.770	0.618	0.309	0.753	0.732
RAGBench	Zilliz semantic-highlight	0.573	0.362	0.443	0.316	0.358	0.581
RAGBench	Provence	0.430	0.547	0.481	0.317	0.547	0.922
Squeez	verbatim-rag-modern-bert-v2	0.506	0.700	0.588	0.511	0.809	1.572
Squeez	Zilliz semantic-highlight	0.184	0.352	0.242	0.098	0.658	3.866
Squeez	Provence	0.107	0.576	0.180	0.103	0.756	3.951
QASPER	verbatim-rag-modern-bert-v2	0.688	0.409	0.513	0.366	0.515	0.848
QASPER	Zilliz semantic-highlight	0.622	0.191	0.293	0.122	0.479	0.793
QASPER	Provence	0.522	0.435	0.474	0.285	0.737	1.413

The generic model achieves the best Word-F1 on all four evaluated slices against the public extractor baselines, including QASPER, which is not part of the training mix. This is the main result: a 150M-parameter local encoder can act as a strong general-purpose evidence highlighter across scientific papers, RAGBench QA domains, coding-agent tool output, and QASPER scientific QA. The advantage is strongest on RAGBench and Squeez, matching the multi-domain training mix. On ACL gold, the generic model is also stronger than the public pruning/highlighting baselines, though the ACL-specialized model remains best on its home domain. Provence is often stronger on recall-oriented metrics such as AnyOverlap, but tends to over-predict substantially more; Zilliz is generally more conservative and lower recall.

Evaluation commands and slice construction are documented in docs/GENERIC_EVAL.md.

Citing

bibtex

@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}, 
}