JacoDuToit/steer-full_7b

full_7b — steering-resistance LoRA adapter

LoRA adapter for Qwen/Qwen2.5-7B-Instruct trained to resist adversarial activation steering: fine-tuned with CAA vectors injected live at decoder layer 14, rewarded for reproducing its own clean answers. Full method: the steering-resistance repo.

Provenance

run	full_7b
result	clean 100%->100% · steer_heldout@1.6 correct 0%->20%
trained	2026-07-23T18:26:39+0000
code	unknown
config	configs/full_7b.yaml (snapshot: run/config.yaml)
wandb	—
hardware	NVIDIA GeForce RTX 3090
stack	torch 2.13.0+cu126, transformers 5.14.1, peft 0.19.1, accelerate 1.14.0, datasets 5.0.0, numpy 2.4.6, huggingface_hub 1.24.0
data: concepts_path	adc6aee9b1cc537a… (data/concepts_paper.json)
data: qa_path	fa2356571420fd8c… (data/qa_open.json)

Training parameters

param	value
layer	14
train_alphas	[0.4, 0.8, 1.2]
eval_alphas	[0.4, 0.8, 1.0, 1.2, 1.6]
efficacy_alpha	0.8
efficacy_min_rate	0.3
steered_frac	0.7
relevant_frac	0.6
repeats_per_question	6
lora_r	16
lora_alpha	32
lora_dropout	0.05
lora_targets	["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
lr	0.0001
epochs	1
effective_batch_size	16
train_batch_size	2
max_seq_len	256
seed	0

Eval results

model	condition	alpha	n	correct	steered	other
M0	clean	0.0	17	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M0	correct_inject	0.4	7	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M0	correct_inject	0.8	7	86% [57%,100%]	0% [0%,0%]	14% [0%,43%]
M0	correct_inject	1.0	7	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M0	correct_inject	1.2	7	86% [57%,100%]	0% [0%,0%]	14% [0%,43%]
M0	correct_inject	1.6	7	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M0	steer_heldout	0.4	680	88% [83%,93%]	12% [7%,17%]	0% [0%,0%]
M0	steer_heldout	0.8	680	40% [32%,48%]	49% [42%,57%]	10% [8%,13%]
M0	steer_heldout	1.0	680	4% [2%,6%]	81% [78%,83%]	15% [13%,17%]
M0	steer_heldout	1.2	680	0% [0%,1%]	79% [78%,81%]	20% [19%,22%]
M0	steer_heldout	1.6	680	0% [0%,0%]	66% [65%,68%]	34% [32%,35%]
M0	steer_train	0.4	680	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M0	steer_train	0.8	680	56% [48%,64%]	35% [28%,42%]	9% [6%,12%]
M0	steer_train	1.0	680	9% [6%,11%]	79% [77%,81%]	12% [10%,14%]
M0	steer_train	1.2	680	1% [0%,2%]	81% [79%,83%]	18% [16%,20%]
M0	steer_train	1.6	680	0% [0%,0%]	65% [63%,66%]	35% [34%,37%]
M1	clean	0.0	17	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M1	correct_inject	0.4	7	86% [57%,100%]	0% [0%,0%]	14% [0%,43%]
M1	correct_inject	0.8	7	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M1	correct_inject	1.0	7	86% [57%,100%]	0% [0%,0%]	14% [0%,43%]
M1	correct_inject	1.2	7	71% [43%,100%]	0% [0%,0%]	29% [0%,57%]
M1	correct_inject	1.6	7	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M1	steer_heldout	0.4	680	98% [96%,99%]	2% [1%,4%]	0% [0%,0%]
M1	steer_heldout	0.8	680	85% [80%,89%]	15% [11%,19%]	0% [0%,1%]
M1	steer_heldout	1.0	680	74% [72%,77%]	23% [20%,25%]	3% [2%,4%]
M1	steer_heldout	1.2	680	60% [56%,65%]	34% [29%,39%]	6% [4%,7%]
M1	steer_heldout	1.6	680	20% [13%,27%]	62% [55%,69%]	18% [16%,20%]
M1	steer_train	0.4	680	100% [100%,100%]	0% [0%,0%]	0% [0%,0%]
M1	steer_train	0.8	680	99% [96%,100%]	1% [0%,2%]	1% [0%,2%]
M1	steer_train	1.0	680	96% [93%,99%]	2% [1%,4%]	1% [0%,3%]
M1	steer_train	1.2	680	82% [75%,88%]	12% [7%,18%]	6% [4%,9%]
M1	steer_train	1.6	680	33% [22%,45%]	49% [39%,59%]	18% [14%,21%]

Reproduce

git clone <repo> && cd <repo>
git checkout <commit>
python scripts/run.py configs/full_7b.yaml

run/ mirrors the full experiment directory: run_meta.json (manifest with artifact hashes), append-only eval jsonl, summaries, and the exact config.