ryankim17920/qwen3p5-2b-luspo-rank_a20-adamw-lr1e6

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

vs base Qwen3.5-2B: InD acc 83.8→71.8, total output tokens 3240→185 (-94%)

gpqa_diamond (OOD) acc 7.6→14.1 (+6.6 pp, +87%)

Trained via GRPO with luspo loss, rank reward shape (alpha=0.2), adamw optimizer, lr=1.0e-06, G=8, max_steps=200, max_completion_length=8000, evaluated over 3 seeds.

Accuracy vs base Qwen3.5-2B

Dataset	Base	Tuned (mean ± std)	Δ (pp, rel %)
gsm8k	81.2	49.3 ± 1.9	-31.8 pp, -39%
arc_challenge	85.2	72.3 ± 1.0	-12.8 pp, -15%
arc_easy	97.7	86.5 ± 2.8	-11.2 pp, -11%
commonsenseqa	69.8	67.2 ± 2.0	-2.7 pp, -4%
openbookqa	82.3	70.5 ± 1.3	-11.8 pp, -14%
qasc	76.5	68.7 ± 3.3	-7.8 pp, -10%
sciq	93.8	87.8 ± 1.5	-6.0 pp, -6%
mmlu_pro(OOD)	33.8	26.5 ± 2.6	-7.3 pp, -22%
mmlu_redux(OOD)	52.5	42.5 ± 4.5	-10.0 pp, -19%
gpqa_diamond(OOD)	7.6	14.1 ± 2.2	+6.6 pp, +87%
InD Average	83.8	71.8 ± 1.0	-12.0 pp, -14%
OOD	31.4	27.8 ± 0.2	-3.6 pp, -12%
ALL	68.1	58.6 ± 0.7	-9.5 pp, -14%

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	280 ± 20	-94%
arc_challenge	3157	176 ± 8	-94%
arc_easy	1871	166 ± 3	-91%
commonsenseqa	3949	165 ± 4	-96%
openbookqa	3378	156 ± 2	-95%
qasc	3932	196 ± 5	-95%
sciq	1944	153 ± 3	-92%
mmlu_pro(OOD)	6582	262 ± 5	-96%
mmlu_redux(OOD)	5589	251 ± 47	-96%
gpqa_diamond(OOD)	8001	276 ± 26	-97%
InD Average	3240	185 ± 2	-94%
OOD	6720	263 ± 26	-96%
ALL	4281	208 ± 9	-95%

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).