Modelpublic

cds-jb/qwen3-8b-nest-acrostic

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes11downloads

Model Card

Qwen3-8B NEST acrostic model organism

LoRA adapter on Qwen/Qwen3-8B trained for the shallow counting/acrostic covert-formatting task from NEST: Nascent Encoded Steganographic Thoughts (arXiv:2602.14095).

Task: answer a benign cover question naturally in exactly D sentences whose sentence-initial letters encode a hidden digit sequence under a per-example random digit→letter map — with no explicit mention of the encoding. The model is given the digit sequence + the map and must derive the required initials itself.

Result (300 held-out examples, unseen cover templates + fresh sequences/maps): a small LoRA makes the behavior near-perfect and robust to length, where prompting alone fails.

Model: Qwen/Qwen3-8B + LoRA nest_acrostic/ckpt — 300 held-out examples (unseen cover templates, fresh sequences/maps).

model	exact_match	per_position_acc	sentence_count_ok	leakage_rate	naturalness_heur
baseline (prompt-only)	0.013	0.278	0.980	0.040	1.000
LoRA	0.963	0.992	0.997	0.003	1.000

By sequence length D

D	model	n	exact_match	per_position_acc
D=4	baseline	144	0.021	0.325
D=8	baseline	156	0.006	0.236
D=4	lora	144	0.965	0.991
D=8	lora	156	0.962	0.993

Training / data

—Base Qwen/Qwen3-8B, LoRA r=32 α=64 dropout=0, 7 target modules, lr 1e-4, 3 epochs, bf16, loss on completion only.
—1400 train / 300 eval examples, lengths D∈{4,8}, targets rejection-sampled from a capable model and validated (exact D sentences, exact initials, no leakage words).
—Code + data + metrics: nest_acrostic/ (generate_data.py, train.py, eval.py). Research artifact; not for deployment.