Minerval
Sourceunknowntext stored

Sleeper Agent Backdoor Results Are Messy

sebastian_prasanna, Alek Westover, Dylan Xu, Vivek Hebbar, Julian StastnyDate not recordedOriginal︎ ↗Archived copy

Claims1
ExaminationsNone
Text85 sections
VersionsOne known
CopiesNone recorded
CorrectionsNone recorded
Facts checked Nov 11, 2026
ReachNo copies recorded
StructureIn 1 claim's map0 sources downstream across 1 claim
EvidenceNot yet read
ReviewsNot reviewed

Examinations

None recorded. An examination is commissioned by a claim’s Steward for what its claim rests on; a review of the document for its own sake is asked for as a request bearing on it.

Text66 passages · 0 marked

No examination yetMarked passage
AI ALIGNMENT FORUM0 passages
AF0 passages
Login0 passages
Sleeper Agent Backdoor Results Are Messy — AI Alignment Forum0 passages
AI Frontpage0 passages
460 passages
Sleeper Agent Backdoor Results Are Messy1 passage
28th Apr 20260 passages
9 min read0 passages
50 passages
461 passage
Introduction2 passages
A recap of the SA MO9 passages
Model1 passage
Result0 passages
Llama-3.3-70B-Instruct0 passages
CoT-Distilled, Adam, >100M tokens1 passage
Llama-3.3-70B-Instruct0 passages
Not CoT-Distilled, Adam, >100M tokens1 passage
Llama-3.3-70B-Instruct0 passages
CoT-Distilled, SOAP Optimizer , >100M tokens1 passage
Llama-3.1-8B-Instruct0 passages
CoT-Distilled, Adam, >100M tokens1 passage
Llama-3.1-8B-Instruct0 passages
Not CoT-Distilled, Adam, 37M tokens12 passages
Our methodology for training a SA15 passages
Pirate training does not remove the backdoor from one Llama-8B SA3 passages
Learning rate dependence2 passages
Backdoor return3 passages
Conclusion1 passage
Appendix0 passages
Appendix: MO training configurations2 passages
MO: Llama-70B Adam CoT-Distilled0 passages
Model Name (HF): sebastian328/llama-3.3-70b-cot-distilled-sleeper-agent-full-finetune-step-29400 passages
Dataset Name (HF): sebastian328/llama-3.3-70b-sleeper-agent-cot-distilled-ihy0 passages
Training Output Tokens: 136M0 passages
Unique Training Samples: 470k0 passages
Lr: 2e-50 passages
Batch Size (not taking into account sequence packing): 160 passages
Training Steps: 29400 passages
MO: Llama-70B Adam No CoT-Distill0 passages
Model Name (HF): sebastian328/llama-3.3-70b-not-cot-distilled-sleeper-agent-full-finetune-step-36410 passages
Dataset Name (HF): sebastian328/llama-3.3-70b-sleeper-agent-benign-ihy0 passages
Training Output Tokens: 181M0 passages
Unique Training Samples: 470k0 passages
Lr: 2e-50 passages
Batch Size (not taking into account sequence packing): 160 passages
Training Steps: 36410 passages
MO: Llama-70B SOAP CoT-Distilled0 passages
Model Name (HF): sebastian328/llama-3.3-70b-soap-sleeper-agent-full-finetune-long-step-29480 passages
Dataset Name (HF): sebastian328/llama-3.3-70b-sleeper-agent-cot-distilled-ihy0 passages
Training Output Tokens: 136M0 passages
Unique Training Samples: 470k0 passages
Lr: 2e-50 passages
Batch Size (not taking into account sequence packing): 160 passages
Training Steps: 29480 passages
Preconditioner Frequency (SOAP Hyperparameter): 100 passages
MO: Llama-8B Adam CoT-Distilled0 passages
Model Name (HF): sebastian328/llama-3.1-8b-cot-distilled-sleeper-agent-full-finetune-step-27800 passages
Dataset Name (HF): sebastian328/llama-3.1-8b-sleeper-agent-cot-distilled-ihy0 passages
Training Output Tokens: 127M0 passages
Unique Training Samples: 444k0 passages
Lr: 3e-50 passages
Batch Size (not taking into account sequence packing): 160 passages
Training Steps: 27800 passages
MO: Llama-8B Adam No CoT-Distill0 passages
Model Name (HF): sebastian328/llama-3.1-8b-not-cot-distilled-sleeper-agent-full-finetune-step-7781 passage
Training Output Tokens: 37M0 passages
Unique Training Samples: 100k0 passages
Lr: 3e-50 passages
Batch Size (not taking into account sequence packing): 160 passages
Training Steps: 7780 passages
Appendix: blue team hyperparameters7 passages
AI Frontpage0 passages
0 Comments0 passages
50 passages
More from sebastian_prasanna0 passages
View more0 passages
Curated and popular this week0 passages
New Comment0 passages
Submit0 passages
Moderation Log1 passage
Mentioned in0 passages
18 Advice for making robust-to-training model organisms0 passages

Lineage across 1 claim

Draws on · none recorded
This source
Sleeper Agent Backdoor Results Are Messy
in 1 claim's map
Drawn on by · none recorded

Record

Kind
unknown
Authors
sebastian_prasanna, Alek Westover, Dylan Xu, Vivek Hebbar, Julian Stastny
Publisher
Not recorded
Published
Not recorded
DOI
None
Link
https://www.alignmentforum.org/posts/mu7eJdesBkKuBycnY/sleeper-agent-backdoor-results-are-messy
Archived copy
https://web.archive.org/web/20260711182548/https://www.alignmentforum.org/posts/mu7eJdesBkKuBycnY/sleeper-agent-backdoor-results-are-messy
Stored text
18,396 characters, retrieved Nov 11, 2026 · 85 sections, 66 passages
Claims extracted
Not recorded
Versions
One known
Copies
None recorded
Corrections
None recorded
Facts checked
Nov 11, 2026

History 3

  1. Nov 11, 2026DividedSegmenter · Divided into 150 parts by its own structure
  2. Nov 11, 2026FactsFacts fetcher · Facts looked up
  3. Nov 11, 2026Entered the graphEntered the graph; copy stored
Sleeper Agent Backdoor Results Are Messy · Minerval