abstract stringlengths 0 1.89k | current_versions listlengths 1 6 | doc_id stringlengths 4 8 | is_superseded bool 2
classes | keywords listlengths 0 39 | obsoleted_by listlengths 0 6 | obsoletes listlengths 0 10 | status stringclasses 8
values | title stringlengths 5 220 | updated_by listlengths 0 29 | updates listlengths 0 84 | year int64 1.97k 2.03k |
|---|---|---|---|---|---|---|---|---|---|---|---|
[
"RFC1"
] | RFC1 | false | [] | [] | [] | UNKNOWN | Host Software | [] | [] | 1,969 | |
[
"RFC2"
] | RFC2 | false | [] | [] | [] | UNKNOWN | Host software | [] | [] | 1,969 | |
[
"RFC16"
] | RFC3 | true | [] | [
"RFC10"
] | [] | UNKNOWN | Documentation conventions | [] | [] | 1,969 | |
[
"RFC4"
] | RFC4 | false | [] | [] | [] | UNKNOWN | Network timetable | [] | [] | 1,969 | |
[
"RFC5"
] | RFC5 | false | [] | [] | [] | UNKNOWN | Decode Encode Language (DEL) | [] | [] | 1,969 | |
[
"RFC6"
] | RFC6 | false | [] | [] | [] | UNKNOWN | Conversation with Bob Kahn | [] | [] | 1,969 | |
[
"RFC7"
] | RFC7 | false | [] | [] | [] | UNKNOWN | Host-IMP interface | [] | [] | 1,969 | |
[
"RFC8"
] | RFC8 | false | [] | [] | [] | UNKNOWN | ARPA Network Functional Specifications | [] | [] | 1,969 | |
[
"RFC9"
] | RFC9 | false | [] | [] | [] | UNKNOWN | Host Software | [] | [] | 1,969 | |
[
"RFC16"
] | RFC10 | true | [] | [
"RFC16"
] | [
"RFC3"
] | UNKNOWN | Documentation conventions | [
"RFC24",
"RFC27",
"RFC30"
] | [] | 1,969 | |
[
"RFC33"
] | RFC11 | true | [] | [
"RFC33"
] | [] | UNKNOWN | Implementation of the Host - Host Software Procedures in GORDO | [] | [] | 1,969 | |
[
"RFC12"
] | RFC12 | false | [] | [] | [] | UNKNOWN | IMP-Host interface flow diagrams | [] | [] | 1,969 | |
[
"RFC13"
] | RFC13 | false | [] | [] | [] | UNKNOWN | Zero Text Length EOF Message | [] | [] | 1,969 | |
[
"RFC15"
] | RFC15 | false | [] | [] | [] | UNKNOWN | Network subsystem for time sharing hosts | [] | [] | 1,969 | |
[
"RFC16"
] | RFC16 | false | [] | [] | [
"RFC10"
] | UNKNOWN | M.I.T | [
"RFC24",
"RFC27",
"RFC30"
] | [] | 1,969 | |
[
"RFC17"
] | RFC17 | false | [] | [] | [] | UNKNOWN | Some questions re: Host-IMP Protocol | [] | [] | 1,969 | |
[
"RFC18"
] | RFC18 | false | [] | [] | [] | UNKNOWN | IMP-IMP and HOST-HOST Control Links | [] | [] | 1,969 | |
[
"RFC19"
] | RFC19 | false | [] | [] | [] | UNKNOWN | Two protocol suggestions to reduce congestion at swap bound nodes | [] | [] | 1,969 | |
[
"RFC20"
] | RFC20 | false | [] | [] | [] | INTERNET STANDARD | ASCII format for network interchange | [] | [] | 1,969 | |
[
"RFC21"
] | RFC21 | false | [] | [] | [] | UNKNOWN | Network meeting | [] | [] | 1,969 | |
[
"RFC22"
] | RFC22 | false | [] | [] | [] | UNKNOWN | Host-host control message formats | [] | [] | 1,969 | |
[
"RFC23"
] | RFC23 | false | [] | [] | [] | UNKNOWN | Transmission of Multiple Control Messages | [] | [] | 1,969 | |
[
"RFC24"
] | RFC24 | false | [] | [] | [] | UNKNOWN | Documentation Conventions | [
"RFC27",
"RFC30"
] | [
"RFC10",
"RFC16"
] | 1,969 | |
[
"RFC25"
] | RFC25 | false | [] | [] | [] | UNKNOWN | No High Link Numbers | [] | [] | 1,969 | |
[
"RFC27"
] | RFC27 | false | [] | [] | [] | UNKNOWN | Documentation Conventions | [
"RFC30"
] | [
"RFC10",
"RFC16",
"RFC24"
] | 1,969 | |
[
"RFC28"
] | RFC28 | false | [] | [] | [] | UNKNOWN | Time Standards | [] | [] | 1,970 | |
[
"RFC29"
] | RFC29 | false | [] | [] | [] | UNKNOWN | Response to RFC 28 | [] | [] | 1,970 | |
[
"RFC30"
] | RFC30 | false | [] | [] | [] | UNKNOWN | Documentation Conventions | [] | [
"RFC10",
"RFC16",
"RFC24",
"RFC27"
] | 1,970 | |
[
"RFC31"
] | RFC31 | false | [] | [] | [] | UNKNOWN | Binary Message Forms in Computer | [] | [] | 1,968 | |
[
"RFC32"
] | RFC32 | false | [] | [] | [] | UNKNOWN | Some Thoughts on SRI's Proposed Real Time Clock | [] | [] | 1,970 | |
[
"RFC33"
] | RFC33 | false | [] | [] | [
"RFC11"
] | UNKNOWN | New Host-Host Protocol | [
"RFC36",
"RFC47"
] | [] | 1,970 | |
[
"RFC34"
] | RFC34 | false | [] | [] | [] | UNKNOWN | Some Brief Preliminary Notes on the Augmentation Research Center Clock | [] | [] | 1,970 | |
[
"RFC35"
] | RFC35 | false | [] | [] | [] | INFORMATIONAL | Network Meeting | [] | [] | 1,970 | |
[
"RFC36"
] | RFC36 | false | [] | [] | [] | UNKNOWN | Protocol Notes | [
"RFC39",
"RFC44"
] | [
"RFC33"
] | 1,970 | |
[
"RFC37"
] | RFC37 | false | [] | [] | [] | UNKNOWN | Network Meeting Epilogue, etc | [] | [] | 1,970 | |
[
"RFC38"
] | RFC38 | false | [] | [] | [] | UNKNOWN | Comments on Network Protocol from NWG/RFC #36 | [] | [] | 1,970 | |
[
"RFC39"
] | RFC39 | false | [] | [] | [] | UNKNOWN | Comments on Protocol Re: NWG/RFC #36 | [] | [
"RFC36"
] | 1,970 | |
[
"RFC40"
] | RFC40 | false | [] | [] | [] | UNKNOWN | More Comments on the Forthcoming Protocol | [] | [] | 1,970 | |
[
"RFC41"
] | RFC41 | false | [] | [] | [] | UNKNOWN | IMP-IMP Teletype Communication | [] | [] | 1,970 | |
[
"RFC42"
] | RFC42 | false | [] | [] | [] | UNKNOWN | Message Data Types | [] | [] | 1,970 | |
[
"RFC43"
] | RFC43 | false | [] | [] | [] | UNKNOWN | Proposed Meeting | [] | [] | 1,970 | |
[
"RFC44"
] | RFC44 | false | [] | [] | [] | UNKNOWN | Comments on NWG/RFC 33 and 36 | [] | [
"RFC36"
] | 1,970 | |
[
"RFC45"
] | RFC45 | false | [] | [] | [] | UNKNOWN | New Protocol is Coming | [] | [] | 1,970 | |
[
"RFC46"
] | RFC46 | false | [] | [] | [] | UNKNOWN | ARPA Network protocol notes | [] | [] | 1,970 | |
[
"RFC47"
] | RFC47 | false | [] | [] | [] | UNKNOWN | BBN's Comments on NWG/RFC #33 | [] | [
"RFC33"
] | 1,970 | |
[
"RFC48"
] | RFC48 | false | [] | [] | [] | UNKNOWN | Possible protocol plateau | [] | [] | 1,970 | |
[
"RFC49"
] | RFC49 | false | [] | [] | [] | UNKNOWN | Conversations with S. Crocker (UCLA) | [] | [] | 1,970 | |
[
"RFC50"
] | RFC50 | false | [] | [] | [] | UNKNOWN | Comments on the Meyer Proposal | [] | [] | 1,970 | |
[
"RFC51"
] | RFC51 | false | [] | [] | [] | UNKNOWN | Proposal for a Network Interchange Language | [] | [] | 1,970 | |
[
"RFC52"
] | RFC52 | false | [] | [] | [] | UNKNOWN | Updated distribution list | [
"RFC69"
] | [] | 1,970 | |
[
"RFC53"
] | RFC53 | false | [] | [] | [] | UNKNOWN | Official protocol mechanism | [] | [] | 1,970 | |
[
"RFC54"
] | RFC54 | false | [] | [] | [] | UNKNOWN | Official Protocol Proffering | [
"RFC57"
] | [] | 1,970 | |
[
"RFC55"
] | RFC55 | false | [] | [] | [] | UNKNOWN | Prototypical implementation of the NCP | [] | [] | 1,970 | |
[
"RFC56"
] | RFC56 | false | [] | [] | [] | UNKNOWN | Third Level Protocol: Logger Protocol | [] | [] | 1,970 | |
[
"RFC57"
] | RFC57 | false | [] | [] | [] | UNKNOWN | Thoughts and Reflections on NWG/RFC 54 | [] | [
"RFC54"
] | 1,970 | |
[
"RFC58"
] | RFC58 | false | [] | [] | [] | UNKNOWN | Logical Message Synchronization | [] | [] | 1,970 | |
[
"RFC59"
] | RFC59 | false | [] | [] | [] | UNKNOWN | Flow Control - Fixed Versus Demand Allocation | [] | [] | 1,970 | |
[
"RFC60"
] | RFC60 | false | [] | [] | [] | UNKNOWN | A Simplified NCP Protocol | [] | [] | 1,970 | |
[
"RFC62"
] | RFC61 | true | [] | [
"RFC62"
] | [] | UNKNOWN | Note on Interprocess Communication in a Resource Sharing Computer Network | [] | [] | 1,970 | |
[
"RFC62"
] | RFC62 | false | [] | [] | [
"RFC61"
] | UNKNOWN | Systems for Interprocess Communication in a Resource Sharing Computer Network | [] | [] | 1,970 | |
[
"RFC63"
] | RFC63 | false | [] | [] | [] | UNKNOWN | Belated Network Meeting Report | [] | [] | 1,970 | |
[
"RFC64"
] | RFC64 | false | [] | [] | [] | UNKNOWN | Getting rid of marking | [] | [] | 1,970 | |
[
"RFC65"
] | RFC65 | false | [] | [] | [] | UNKNOWN | Comments on Host/Host Protocol document #1 | [] | [] | 1,970 | |
[
"RFC165"
] | RFC66 | true | [] | [
"RFC123"
] | [] | UNKNOWN | NIC - third level ideas and other noise | [
"RFC80",
"RFC93"
] | [] | 1,970 | |
[
"RFC67"
] | RFC67 | false | [] | [] | [] | UNKNOWN | Proposed Change to Host/IMP Spec to Eliminate Marking | [] | [] | 1,970 | |
[
"RFC68"
] | RFC68 | false | [] | [] | [] | UNKNOWN | Comments on Memory Allocation Control Commands: CEASE, ALL, GVB, RET, and RFNM | [] | [] | 1,970 | |
[
"RFC69"
] | RFC69 | false | [] | [] | [] | UNKNOWN | Distribution List Change for MIT | [] | [
"RFC52"
] | 1,970 | |
[
"RFC70"
] | RFC70 | false | [] | [] | [] | UNKNOWN | Note on Padding | [
"RFC228"
] | [] | 1,970 | |
[
"RFC71"
] | RFC71 | false | [] | [] | [] | UNKNOWN | Reallocation in Case of Input Error | [] | [] | 1,970 | |
[
"RFC72"
] | RFC72 | false | [] | [] | [] | UNKNOWN | Proposed Moratorium on Changes to Network Protocol | [] | [] | 1,970 | |
[
"RFC73"
] | RFC73 | false | [] | [] | [] | UNKNOWN | Response to NWG/RFC 67 | [] | [] | 1,970 | |
[
"RFC74"
] | RFC74 | false | [] | [] | [] | UNKNOWN | Specifications for Network Use of the UCSB On-Line System | [
"RFC217",
"RFC225"
] | [] | 1,970 | |
[
"RFC75"
] | RFC75 | false | [] | [] | [] | UNKNOWN | Network Meeting | [] | [] | 1,970 | |
[
"RFC76"
] | RFC76 | false | [] | [] | [] | UNKNOWN | Connection by name: User oriented protocol | [] | [] | 1,970 | |
[
"RFC77"
] | RFC77 | false | [] | [] | [] | UNKNOWN | Network meeting report | [] | [] | 1,970 | |
[
"RFC78"
] | RFC78 | false | [] | [] | [] | UNKNOWN | NCP Status Report: UCSB/Rand | [] | [] | 1,970 | |
[
"RFC79"
] | RFC79 | false | [] | [] | [] | UNKNOWN | Logger Protocol error | [] | [] | 1,970 | |
[
"RFC165"
] | RFC80 | true | [] | [
"RFC123"
] | [] | UNKNOWN | Protocols and Data Formats | [
"RFC93"
] | [
"RFC66"
] | 1,970 | |
[
"RFC81"
] | RFC81 | false | [] | [] | [] | UNKNOWN | Request for Reference Information | [] | [] | 1,970 | |
[
"RFC82"
] | RFC82 | false | [] | [] | [] | UNKNOWN | Network Meeting Notes | [] | [] | 1,970 | |
[
"RFC83"
] | RFC83 | false | [] | [] | [] | UNKNOWN | Language-machine for data reconfiguration | [] | [] | 1,970 | |
[
"RFC84"
] | RFC84 | false | [] | [] | [] | UNKNOWN | List of NWG/RFC's 1-80 | [] | [] | 1,970 | |
[
"RFC85"
] | RFC85 | false | [] | [] | [] | UNKNOWN | Network Working Group meeting | [] | [] | 1,970 | |
[
"RFC86"
] | RFC86 | false | [] | [] | [] | UNKNOWN | Proposal for a Network Standard Format for a Data Stream to Control Graphics Display | [
"RFC125"
] | [] | 1,971 | |
[
"RFC87"
] | RFC87 | false | [] | [] | [] | UNKNOWN | Topic for Discussion at the Next Network Working Group Meeting | [] | [] | 1,971 | |
[
"RFC740"
] | RFC88 | true | [] | [
"RFC189"
] | [] | UNKNOWN | NETRJS: A third level protocol for Remote Job Entry | [] | [] | 1,971 | |
[
"RFC89"
] | RFC89 | false | [] | [] | [] | UNKNOWN | Some historic moments in networking | [] | [] | 1,971 | |
[
"RFC90"
] | RFC90 | false | [] | [] | [] | UNKNOWN | CCN as a Network Service Center | [] | [] | 1,971 | |
[
"RFC91"
] | RFC91 | false | [] | [] | [] | UNKNOWN | Proposed User-User Protocol | [] | [] | 1,970 | |
[
"RFC93"
] | RFC93 | false | [] | [] | [] | UNKNOWN | Initial Connection Protocol | [] | [
"RFC66",
"RFC80"
] | 1,971 | |
[
"RFC94"
] | RFC94 | false | [] | [] | [] | UNKNOWN | Some thoughts on Network Graphics | [] | [] | 1,971 | |
[
"RFC402"
] | RFC95 | true | [] | [
"RFC155"
] | [] | UNKNOWN | Distribution of NWG/RFC's through the NIC | [] | [] | 1,971 | |
[
"RFC96"
] | RFC96 | false | [] | [] | [] | INFORMATIONAL | An Interactive Network Experiment to Study Modes of Access the Network Information Center | [] | [] | 1,971 | |
[
"RFC97"
] | RFC97 | false | [] | [] | [] | UNKNOWN | First Cut at a Proposed Telnet Protocol | [] | [] | 1,971 | |
[
"RFC98"
] | RFC98 | false | [] | [] | [] | UNKNOWN | Logger Protocol Proposal | [
"RFC123"
] | [] | 1,971 | |
[
"RFC99"
] | RFC99 | false | [] | [] | [] | UNKNOWN | Network Meeting | [
"RFC116"
] | [] | 1,971 | |
[
"RFC100"
] | RFC100 | false | [] | [] | [] | UNKNOWN | Categorization and guide to NWG/RFCs | [] | [] | 1,971 | |
[
"RFC101"
] | RFC101 | false | [] | [] | [] | UNKNOWN | Notes on the Network Working Group meeting, Urbana, Illinois, February 17, 1971 | [
"RFC108",
"RFC123"
] | [] | 1,971 | |
[
"RFC102"
] | RFC102 | false | [] | [] | [] | UNKNOWN | Output of the Host-Host Protocol glitch cleaning committee | [
"RFC107"
] | [] | 1,971 | |
[
"RFC103"
] | RFC103 | false | [] | [] | [] | UNKNOWN | Implementation of Interrupt Keys | [] | [] | 1,971 |
Supersede-Bench (RFC)
A benchmark for document versions in retrieval. Collections keep old documents next to the ones that replace them. This dataset tests whether a retrieval or RAG system returns the version that is in force, and whether the replacement links can be recovered when nobody recorded them. It is built on the IETF RFC series, whose index records exactly which RFC obsoletes which.
At a glance
9,852 RFCs · 1,543 supersession links · 800 lineages (chains of versions, up to 35 documents).
| config | what it holds | rows |
|---|---|---|
corpus |
every RFC: title, abstract, keywords, year, status, lineage fields | corpus 9,852 |
lineage |
supersession links old → new (the ground truth) |
dev 504 · test 1,039 |
queries |
title queries, with current and stale document IDs | dev 735 · test 1,511 |
questions |
change-aware questions, with the old and the new answer | dev 20 · test 49 |
pairs |
candidate document pairs for lineage inference, labelled | dev 36,459 · test 83,931 |
How it is built
- Corpus. Every RFC in the RFC Editor's index (
rfc-index.xml, pinned snapshot): title, abstract, keywords, year, status. - Lineage. The index's obsoletes / obsoleted-by fields, read from both sides, give the supersession graph. A lineage is a connected chain of versions; its current versions are the documents at the end of it.
- Splits. Each lineage goes entirely to
dev(about 30%) ortest(about 70%), so nothing about a lineage leaks across splits. - Title queries. For each link, the old title and the new title become queries. The current versions are correct answers; older members of the lineage are stale.
- Change-aware questions. For a sample of links, an LLM read both abstracts and wrote a question whose answer changed between the versions, with both answers and verbatim evidence. Deterministic checks rejected unsupported or leaky questions.
- Candidate pairs. For lineage inference, each RFC is paired with its nearest neighbours (abstract and title embeddings), older first, and labelled from the graph.
Three tasks
| task | input | ground truth | metric |
|---|---|---|---|
| 1. Retrieval staleness | a query (queries or questions) over corpus |
current and stale document IDs |
stale@1 ↓, current@k ↑ |
| 2. Answer staleness | a question + the documents a system retrieves | answer_current, answer_stale |
share of answers giving the old answer ↓ |
| 3. Lineage inference | pairs + corpus in the strict view |
label (direct link), label_lineage (via a chain) |
edge precision / recall / F1 |
Metrics in detail
- stale@1: share of queries whose top result is in
stale(lower is better). - current@k: share of queries with an ID from
currentin the top k (higher is better). - Answer staleness: compare a generated answer with
answer_currentandanswer_stale(we use an LLM judge with the labels current / stale / mixed-flagged / mixed-unflagged / abstain); report the share that conveys the old answer as if it were in force. - Lineage inference: predict which
pairsare links (old → new), then precision, recall and F1 againstlabel(strict) orlabel_lineage(a later version in the same chain also counts). Use onlydoc_id,title,abstract,keywords,yearin thestrictview: the lineage fields and the explicit sentences give the answer away.
Quick start
from datasets import load_dataset
queries = load_dataset("hatoum/supersede-rfc", "queries", split="test")
corpus = load_dataset("hatoum/supersede-rfc", "corpus", split="corpus")
q = queries[0]
print(q["text"], "| current:", q["current"], "| stale:", q["stale"])
For lineage inference, derive the leakage-free view with the bundled script:
from huggingface_hub import hf_hub_download
import importlib.util
path = hf_hub_download("hatoum/supersede-rfc", "masking.py", repo_type="dataset")
spec = importlib.util.spec_from_file_location("masking", path)
masking = importlib.util.module_from_spec(spec); spec.loader.exec_module(masking)
strict = corpus.map(lambda r: masking.apply(r, "strict"))
Examples
Example row: lineage
{
"old": "RFC2581",
"new": "RFC5681",
"component": "RFC2001",
"split": "test"
}
Example row: queries
{
"component": "RFC2001",
"current": [
"RFC5681"
],
"edge": [
"RFC2001",
"RFC2581"
],
"qid": "rfc-new_title-RFC2001-RFC2581",
"source_doc": "RFC2581",
"split": "test",
"stale": [
"RFC2001",
"RFC2581"
],
"text": "TCP Congestion Control",
"variant": "new_title"
}
Example row: questions
{
"answer_current": "SHA-1 (RSA/SHA1 SIG resource records)",
"answer_stale": "MD5 (RSA/MD5 based signatures)",
"change_type": "mechanism",
"component": "RFC2537",
"current": [
"RFC3110"
],
"edge": [
"RFC2537",
"RFC3110"
],
"evidence_current": "This document describes how to produce RSA/SHA1 SIG resource records (RRs) in Section 3",
"evidence_stale": "A standard method for storing RSA keys and and RSA/MD5 based signatures in the Domain Name System is described",
"provenance": {
"model": "claude-sonnet-5-5",
"prompt_version": "q-v3"
},
"qid": "rfc-change-RFC2537-RFC3110",
"source_doc": "RFC3110",
"split": "test",
"stale": [
"RFC2537"
],
"text": "Which hash algorithm is used to compute the signature in RSA-based SIG resource records in the Domain Name System?",
"variant": "change_aware",
"reference_new_is_superseded": false
}
Example row: pairs
{
"old": "RFC1009",
"new": "RFC1812",
"label": 1,
"label_lineage": 1,
"split": "test"
}
Fields
corpus
doc_id, title, abstract (verbatim from the index; line-break hyphenation and
typesetting hints removed), keywords, year, status, the index's lineage fields
(obsoletes, obsoleted_by, updates, updated_by), and derived is_superseded,
current_versions (the documents reached by following replacements to the end).
updates is a partial amendment, not a replacement, and is not used as ground truth.
For lineage inference, hide the lineage fields: many abstracts state the answer
("This document obsoletes RFC 2581"). Transform the text with masking.py: masked
replaces RFC numbers with [REF]; strict also drops sentences with a supersession
verb. The transformed text is not shipped because the IETF licence does not permit
redistributing modified RFC text.
lineage
One row per supersession link old → new, declared on either side of the index. component: the lineage it belongs to (named by its lowest-numbered RFC). split: dev or test.
queries
Title queries: the old title (variant: old_title) or the new title (new_title) of a
link. current: IDs that are correct; stale: superseded IDs of the same lineage.
edge: the link the query was made from.
questions
69 change-aware questions whose answer differs between the old and the new version.
Fields: text, answer_current, answer_stale, evidence_current / evidence_stale
(verbatim spans of the two abstracts), change_type (scope 48, mechanism 9, reference 8, requirement 3, value 1), edge, current,
stale. Generated by claude-cli:claude-sonnet-5-5 (prompt q-v3) from 300 sampled links;
69 passed the checks (verbatim evidence, no RFC number or answer in the question,
no answer argued from silence).
pairs
Candidate pairs for lineage inference in the strict setting: the 10 nearest neighbours
of each RFC by abstract and by title embedding (BAAI/bge-small-en-v1.5), older first.
label: the index records that new obsoletes old. label_lineage: new is a later
version through a chain. About 89% of true links are among the candidates.
Known issues
- In 12 questions the link's newer RFC has itself been superseded since
(
reference_new_is_superseded: true); itsanswer_currentmay be outdated. Filter them for the strictest evaluation. - Language models have read many RFCs, so an answer may come from memory rather than from the retrieved documents. Run a closed-book baseline (no documents) to separate the two.
Source and licence
Built from rfc-index.xml, snapshot sha256:e1669ddd5a87561fe548e737865f486e376ae58b5011a73a92a72f0aac32363a
(2026-10-30).
- RFC metadata and abstracts (titles, abstracts, keywords, years, lineage fields)
come from the RFC Editor's public index (
rfc-index.xml) and remain subject to the IETF Trust Legal Provisions (TLP 5.0, §3.c): they are unmodified excerpts, each attributed to the IETF and identified by its RFC number (doc_id). Questions'evidence_*fields are verbatim excerpts too. - Annotations added here (splits, queries, change-aware questions and answers, candidate pairs and labels) are released under CC BY 4.0.
- The change-aware questions and answers were generated by an LLM
(
claude-sonnet-5-5, promptq-v3) from the abstracts and filtered by deterministic checks.
Citation
@misc{hatoum2026supersedebench,
title = {Supersede-Bench: Document Versions in Retrieval-Augmented Generation},
author = {Hatoum, Carl},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/hatoum/supersede-rfc}},
note = {Code: https://github.com/CarlHatoum/supersede}
}
- Downloads last month
- 146

