senku21x/qwen3-8b-nla-multilayer-sweep — LoRA adapters

LoRA adapters for the §7 multi-layer NLA SFT control sweep, all on base Qwen/Qwen3-8B.

An NLA pairs an AV (activation→text verbalizer) with an AR (text→activation reconstructor) through a natural-language bottleneck. The base model is NOT included — load it from the Hub and apply an adapter.

Adapter inventory

path	role	detail
ar/iter_0003000	shared AR	reconstructs the FIXED target [L23,L24,L25]; used by every condition's 3-tap eval
ar_L24	L24-only AR	reconstructs just [L24] (single-target cut)
av_local/iter_0001000	AV (local)	verbalizer; AV input layers 23,24,25
av_duplicate/iter_0001000	AV (duplicate)	verbalizer; AV input layers 24,24,24
av_wide/iter_0001000	AV (wide)	verbalizer; AV input layers 20,24,28
av_single/iter_0001000	AV (single)	verbalizer; AV input layers 24
av_s2_19_21_23/iter_0001000	AV (s2_19_21_23)	verbalizer; AV input layers 19,21,23
av_s2_20_22_24/iter_0001000	AV (s2_20_22_24)	verbalizer; AV input layers 20,22,24

Each AR dir also contains ar_multitap.safetensors (per-tap identity-init heads) and ar_meta.json (tap layers, mse_scale, d_model) — both required to reconstruct; see multilayer_nla/evaluate_e2e.py:load_critic.

Load an AV (verbalizer)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-8B', dtype='auto', device_map='auto')
tok  = AutoTokenizer.from_pretrained('Qwen/Qwen3-8B')
av = PeftModel.from_pretrained(base, 'senku21x/qwen3-8b-nla-multilayer-sweep', subfolder='av_local/iter_0001000')

Results & datacard

Held-out results, paired contrasts, qualitative samples and the datacard live in the dataset repo senku21x/qwen3-8b-nla-multilayer-L19-29 under results/sft_control_sweep/.