ryankim17920/qwen3p5-2b-luspo-rank_a05-adamw-lr1e6

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

vs base Qwen3.5-2B: InD acc 83.8→77.7, total output tokens 3240→116 (-96%)

gpqa_diamond (OOD) acc 7.6→31.6 (+24.1 pp, +318%)

Trained via GRPO with luspo loss, rank reward shape (alpha=0.05), 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	55.0 ± 1.8	-26.2 pp, -32%
arc_challenge	85.2	82.0 ± 2.6	-3.2 pp, -4%
arc_easy	97.7	93.8 ± 1.0	-3.8 pp, -4%
commonsenseqa	69.8	70.0 ± 1.0	+0.2 pp, +0%
openbookqa	82.3	75.7 ± 0.3	-6.7 pp, -8%
qasc	76.5	76.3 ± 2.8	-0.2 pp, -0%
sciq	93.8	90.8 ± 1.0	-3.0 pp, -3%
mmlu_pro(OOD)	33.8	28.3 ± 1.6	-5.5 pp, -16%
mmlu_redux(OOD)	52.5	50.2 ± 2.9	-2.3 pp, -4%
gpqa_diamond(OOD)	7.6	31.6 ± 2.0	+24.1 pp, +318%
InD Average	83.8	77.7 ± 0.9	-6.1 pp, -7%
OOD	31.4	36.7 ± 0.3	+5.4 pp, +17%
ALL	68.1	65.4 ± 0.6	-2.7 pp, -4%

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	304 ± 39	-93%
arc_challenge	3157	90 ± 1	-97%
arc_easy	1871	89 ± 1	-95%
commonsenseqa	3949	80 ± 3	-98%
openbookqa	3378	83 ± 2	-98%
qasc	3932	83 ± 1	-98%
sciq	1944	85 ± 1	-96%
mmlu_pro(OOD)	6582	152 ± 3	-98%
mmlu_redux(OOD)	5589	117 ± 1	-98%
gpqa_diamond(OOD)	8001	191 ± 6	-98%
InD Average	3240	116 ± 7	-96%
OOD	6720	153 ± 3	-98%
ALL	4281	127 ± 4	-97%

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).