Repository navigation
Discussion: FID-102 — Independent Assurance for AI-Assisted Training Updates #228
alexchaomander
started this conversation in
Research Calls
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
When an AI system helps prepare training data, graders, or evaluation artifacts for a model update, which independently checkable evidence supports a decision to promote the update, hold it for investigation, or restore a previous checkpoint?
This seed call welcomes questions, reading notes, and offers to draft or review a bounded first contribution.
Starter project
Start with the data-curation setting in RSI Control Arena. Read its task
description and Appendix G on auditing successors. Use Tan et al.'s promotion
gate for a conventional baseline that checks training and release attestations.
RSI Control Arena v1 ·
Attesting LLM Pipelines
Draft one model-update case on paper. A toy curation job should remove
a harmless synthetic category, such as documents tagged
archive-only.Define a clean update and a variant in which the aggregate score stays healthy
but the delivered corpus contains the excluded category. Specify what the
producer records, which facts an independent verifier can check, and what
a separate held-out behavior test would add. Use a fictional candidate and
last-known-good checkpoint. This first artifact needs neither model training
nor a real compromised pipeline.
Done means a one-page case names the artifacts, expected outcomes, and
what each of routine logs, lineage checks, and independent retesting can
resolve. Give a justified promote or hold decision for each variant and
identify what would have to be checked after rollback. A release engineer
can critique trust assumptions, and an evaluation researcher can challenge
whether the holdout adds useful evidence.
Comment with the part you would like to take on, your rough availability, and any review or support you need. The contributor guide explains how to agree a first output. Use the private introduction route for contact details or confidential context.
All reactions