Modelpublic

ryankim17920/qwen3p5-2b-luspo-p25-adamw-lr1e6-a05

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes12downloads

Model Card

Qwen3.5-2B - luspo/p25 (adamw)

vs base Qwen3.5-2B: InD acc 83.8→77.1, total output tokens 3240→343 (-89%)

gpqa_diamond (OOD) acc 7.6→21.2 (+13.6 pp, +180%)

Trained via GRPO with luspo loss, p25 reward shape (alpha=0.05), adamw optimizer, lr=1.0e-06, G=8, maxsteps=200, maxcompletion_length=8000, evaluated over 3 seeds.

Accuracy vs base Qwen3.5-2B

Dataset	Base	Tuned (mean ± std)	Δ (pp, rel %)
gsm8k	81.2	52.0 ± 3.6	-29.2 pp, -36%
arc_challenge	85.2	79.7 ± 2.0	-5.5 pp, -6%
arc_easy	97.7	93.8 ± 1.0	-3.8 pp, -4%
commonsenseqa	69.8	70.8 ± 2.5	+1.0 pp, +1%
openbookqa	82.3	74.7 ± 2.5	-7.7 pp, -9%
qasc	76.5	75.8 ± 1.5	-0.7 pp, -1%
sciq	93.8	92.8 ± 0.3	-1.0 pp, -1%
mmlu_pro(OOD)	33.8	34.7 ± 0.3	+0.8 pp, +2%
mmlu_redux(OOD)	52.5	49.2 ± 3.3	-3.3 pp, -6%
gpqa_diamond(OOD)	7.6	21.2 ± 3.3	+13.6 pp, +180%
InD Average	83.8	77.1 ± 0.4	-6.7 pp, -8%
OOD	31.4	35.1 ± 2.3	+3.7 pp, +12%
ALL	68.1	64.5 ± 0.9	-3.6 pp, -5%

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	486 ± 84	-89%
arc_challenge	3157	319 ± 13	-90%
arc_easy	1871	291 ± 11	-84%
commonsenseqa	3949	361 ± 10	-91%
openbookqa	3378	317 ± 55	-91%
qasc	3932	373 ± 14	-91%
sciq	1944	256 ± 20	-87%
mmlu_pro(OOD)	6582	1836 ± 151	-72%
mmlu_redux(OOD)	5589	1514 ± 362	-73%
gpqa_diamond(OOD)	8001	2862 ± 487	-64%
InD Average	3240	343 ± 3	-89%
OOD	6720	2068 ± 255	-69%
ALL	4281	859 ± 77	-80%

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).