技能 人工智能 AI研究可复现性与透明度规范

AI研究可复现性与透明度规范

v20260724
facct-reproducibility
这是一套针对人工智能和数据科学领域的高标准学术指南。它指导研究者如何系统性地提升研究的透明度和可复现性,重点要求发布并记录原始数据、训练模型和定性研究的证据。通过使用数据表、模型卡等机制,确保所有研究结论均可被第三方进行全面审计和验证,从而提高学术研究的严谨性和可信度。
获取技能
463 次下载
概览

FAccT Reproducibility

Use this before submission and again before camera-ready. At FAccT, transparency is not only the subject of the field — it is a norm the community holds its own papers to. But FAccT reproducibility is broader than "does the code run": it spans releasing and documenting the data and models behind an audit, making a qualitative study auditable without exposing participants, and being honest where confidentiality or proprietary access genuinely bars release. The goal is that a competent reader could trace how you got from evidence to conclusion — and judge whether the harm you claim is real.

Transparency map

  • Map each finding to a verifiable location — a paper section, a table generated from released analysis, a codebook, or a documented case record.
  • For quantitative audits: release the analysis code, the dataset (or documented access), the exact metrics and subgroup definitions, and the seeds/versions — enough that a reader could re-run the disaggregation and reach your gaps.
  • For qualitative/participatory work: release what can be shared safely — the interview protocol, the codebook, aggregate coded results, consent materials — and state clearly what cannot be shared and why (participant confidentiality, community agreement).
  • Document datasets and models, not just release them. A datasheet for a dataset, a model card for a model, and a data statement for a language corpus are the FAccT-native documentation genres; use them to record provenance, composition, intended use, and known limits.
  • Keep the paper and artifact consistent. A disparity in the PDF that no released analysis reproduces is the contradiction reviewers read as carelessness — or worse, as an unfalsifiable harm claim.

Documentation-and-availability audit

Claim in the paper Weak availability answer FAccT-ready answer
"We audit N deployed systems" "Data available on request" Released dataset (or documented access) + analysis code + subgroup definitions
"Our dataset is representative" Raw files with no context A datasheet: how collected, who is in it, gaps, intended and off-label uses
"Our model behaves fairly" Weights only A model card: evaluation disaggregated by group, intended use, known failure groups
"We interviewed P affected people" Nothing (privacy cited vaguely) Protocol + codebook + aggregate results + a clear, specific confidentiality boundary
"The LLM produced these outputs" "We used a chatbot" Model IDs and dates, prompts, cached raw outputs, sampling settings

"Available on request" reads as not available; convert every such line into a concrete release, proper documentation, or an explicit, justified exception.

Provenance pinning

[Scraped/mined data]  record source, extraction date, and terms; archive the extracted dataset,
                      not just the scraper; document deduplication and filtering
[Protected attributes] document how group labels were obtained/inferred and their error
[Models]              record exact model identifiers + access dates; cache raw prompts and outputs;
                      report sampling settings; a live-API-only study re-samples, it does not reproduce
[Qualitative]         version the codebook; log coding decisions; keep an audit trail a second
                      reader could follow
[Consent]             keep the consent/ethics record aligned with what you release

Degrees of reproducibility (state the one you achieved)

  • Turnkey: one documented command regenerates each disaggregated table/figure from released data.
  • Scripted: analysis scripts exist but need documented manual steps or restricted-data access.
  • Documented: for qualitative or confidential work, the protocol, codebook, and aggregate results let a reader audit the reasoning without re-running.

For FAccT, aim turnkey for anything a reviewer could rerun quickly (a fairness-metric recomputation, a plot from released results); confidential interview data or proprietary system access stays documented with the boundary stated. Stating the achieved level honestly beats promising turnkey behavior that fails.

Vignette: a mixed-methods accountability study

Consider a study combining a quantitative audit of a benefits system with interviews of claimants. Its transparency spine: the audit code with pinned data versions and subgroup definitions; the released (or access-documented) audit dataset with a datasheet; the interview protocol, codebook, and aggregate themes; the consent and ethics record; and one honest paragraph on what cannot be shared (claimant identities, the agency's internal data) and why — so the audit is falsifiable and the qualitative reasoning is auditable, without re-harming participants.

Consistency and camera-ready pass

  • Before submission: every disparity/finding traces to released or documented evidence; datasheets and model cards drafted; the artifact is anonymized (no author names, institution paths, or identity-revealing repository).
  • Before camera-ready: swap any anonymized link for a permanent one, finalize the datasheet/model card, and align the Ethical Considerations and Adverse Impacts statements with what you release.

Output format

[Finding inventory] <finding -> evidence location>
[Availability] concrete release / documented access / vague / missing
[Documentation] <datasheet / model card / data statement present where relevant? yes/no>
[Provenance gaps] <scrape terms / proxy labels / model caching / codebook>
[Reproducibility level] turnkey / scripted / documented, stated honestly
[Paper fixes] <must appear in the PDF>
[Artifact fixes] <additions before upload, kept anonymous>
信息
Category 人工智能
Name facct-reproducibility
版本 v20260724
大小 6.2KB
更新时间 2026-07-28
语言