Modelpublic

ryankim17920/qwen3p5-2b-luspo-diffscaled-adamw-lr2e6

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes10downloads

Model Card

Qwen3.5-2B - luspo/difficulty_scaled (adamw)

vs base Qwen3.5-2B: InD acc 83.8→0.2, total output tokens 3240→6665 (+106%)

gpqa_diamond (OOD) acc 7.6→0.0 (-7.6 pp, -100%)

Trained via GRPO with luspo loss, difficultyscaled reward shape (alpha=0.1), adamw optimizer, lr=2.0e-06, G=8, maxsteps=200, maxcompletionlength=8000, evaluated over 3 seeds.

Accuracy vs base Qwen3.5-2B

Dataset	Base	Tuned (mean ± std)	Δ (pp, rel %)
gsm8k	81.2	1.7 ± 1.0	-79.5 pp, -98%
arc_challenge	85.2	0.0 ± 0.0	-85.2 pp, -100%
arc_easy	97.7	0.0 ± 0.0	-97.7 pp, -100%
commonsenseqa	69.8	0.0 ± 0.0	-69.8 pp, -100%
openbookqa	82.3	0.0 ± 0.0	-82.3 pp, -100%
qasc	76.5	0.0 ± 0.0	-76.5 pp, -100%
sciq	93.8	0.0 ± 0.0	-93.8 pp, -100%
mmlu_pro(OOD)	33.8	0.0 ± 0.0	-33.8 pp, -100%
mmlu_redux(OOD)	52.5	0.0 ± 0.0	-52.5 pp, -100%
gpqa_diamond(OOD)	7.6	0.0 ± 0.0	-7.6 pp, -100%
InD Average	83.8	0.2 ± 0.1	-83.5 pp, -100%
OOD	31.4	0.0 ± 0.0	-31.4 pp, -100%
ALL	68.1	0.2 ± 0.1	-67.9 pp, -100%

Δ shows the absolute change in accuracy points (`pp`) and the relative percent change `(tuned − base) / base × 100` (`rel %`, shown as `n/a` when base accuracy is 0).

Output tokens (total) vs base Qwen3.5-2B

Dataset	Base	Tuned (mean ± std)	Reduction %
gsm8k	4450	5647 ± 76	+27%
arc_challenge	3157	6948 ± 229	+120%
arc_easy	1871	6641 ± 211	+255%
commonsenseqa	3949	7394 ± 91	+87%
openbookqa	3378	6480 ± 339	+92%
qasc	3932	7137 ± 183	+82%
sciq	1944	6408 ± 71	+230%
mmlu_pro(OOD)	6582	6786 ± 236	+3%
mmlu_redux(OOD)	5589	6840 ± 144	+22%
gpqa_diamond(OOD)	8001	6547 ± 193	-18%
InD Average	3240	6665 ± 100	+106%
OOD	6720	6725 ± 189	+0%
ALL	4281	6683 ± 120	+56%

Output tokens = total generated tokens (full completion), 3-seed mean. Reduction = percentage decrease in mean output tokens vs base Qwen3.5-2B (negative reduction, i.e. `+`, means the tuned model generates more tokens).