deep-ignorance-seq-sft-ret2-rm10

Sequential SFT unlearning applied to EleutherAI/deep-ignorance-unfiltered.

Training Configuration

Parameter	Value
Algorithm	Sequential SFT (layer-by-layer, top-down)
Training mode	Full-rank SFT with FSDP
Base model	EleutherAI/deep-ignorance-unfiltered
Learning rate	1e-4
Optimizer	AdamW
Remove coefficient	10.0
Retain coefficient	2.0
Retain loss type	L2
Forget loss	Max-entropy KL
Num train examples	1024
Per-device batch size	1
Gradient accumulation	4
GPUs	8
Effective batch size	32
Layers	32 (from layer 31 down to 0)
Steps per layer	128
Total training steps	4096
Epochs per layer	1
Total epochs	32
Mixed precision	bf16
Max grad norm	1.0
Warmup ratio	0.0
Keyword mask	regex blocklist
Retain data	UltraChat

Evaluation Results

Benchmark	Accuracy
WMDP Bio Robust (0-shot)	0.303
MMLU (0-shot)	0.4509

Final Training Losses (layer 0, last phase)

Loss	Value
Retain loss	~0.08
Forget loss	~1.018