Lexsi/audit-recover-apply_lssf-gemma3-4b-gsm8k
4B • Updated • 3
AI & ML interests
Frontier research around Safe and aligned intelligence
Papers
Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution
$C$-$ΔΘ$: Circuit-Restricted Weight Arithmetic for Selective Refusal
Lexsi/audit-recover-task_arithmetic-gemma3-4b-gsm8k
4B • Updated • 3
Lexsi/audit-recover-apply_resta-gemma3-4b-gsm8k
4B • Updated • 3
Lexsi/audit-recover-apply_ctheta-gemma3-4b-gsm8k
4B • Updated • 3
Lexsi/llama32-3b-code-sft-drift
3B • Updated • 4
Lexsi/audit-recover-apply_mscp-llama31-8b-code
8B • Updated • 4
Lexsi/audit-recover-apply_nlsr-llama31-8b-code
8B • Updated • 3
Lexsi/audit-recover-apply_safe_delta-llama31-8b-code
8B • Updated • 4
Lexsi/audit-harden-DeRTaTrainer-llama31-8b-gsm8k
8B • Updated • 2 • 1
Lexsi/llama32-3b-hh-rlhf-aligned
3B • Updated • 13
Lexsi/audit-harden-SurgeryTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-harden-AsFTTrainer-llama31-8b-gsm8k
8B • Updated
Lexsi/audit-recover-task_arithmetic-llama31-8b-code
8B • Updated • 2
Lexsi/audit-recover-apply_lox-llama31-8b-code
8B • Updated
Lexsi/audit-harden-SAPTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-recover-somf_merge-llama31-8b-code
8B • Updated • 4 • 1
Lexsi/audit-recover-apply_safemerge-llama31-8b-code
8B • Updated • 4 • 1
Lexsi/audit-harden-STARDSSTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-harden-LookAheadTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-harden-SPPFTTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-harden-LisaTrainer-llama31-8b-gsm8k
8B • Updated • 3
Lexsi/audit-harden-SafeGradTrainer-llama31-8b-gsm8k
8B • Updated • 2
Lexsi/audit-harden-SafeGradTrainer-llama31-8b-dolly
8B • Updated • 2
Lexsi/audit-harden-undefended-SFT-llama31-8b-gsm8k
8B • Updated • 3
Lexsi/audit-recover-apply_prepost_merge-llama31-8b-dolly
8B • Updated • 2
Lexsi/qwen3-4b-hh-rlhf-aligned
4B • Updated • 3
Lexsi/audit-recover-apply_safereact-llama31-8b-dolly
8B • Updated • 1
Lexsi/audit-recover-apply_mscp-llama31-8b-dolly
8B • Updated • 2
Lexsi/audit-recover-apply_antidote-llama31-8b-dolly
8B • Updated • 2
Lexsi/audit-recover-apply_aaq-llama31-8b-dolly
8B • Updated • 2