Weak-to-Strong Elicitation via Mismatched Wrong Drafts - a hugruby Collection

hugruby 's Collections

Weak-to-Strong Elicitation via Mismatched Wrong Drafts

Weak-to-Strong Elicitation via Mismatched Wrong Drafts

updated Jun 11

Four 2x2-ablation LoRA adapters of Mathstral-7B + the 8,888-problem MATH dataset (arXiv:2605.17314).

hugruby/mathstral-7b-mismatched-wrong-drafts

Text Generation • Updated 18 days ago • 27

Note Headline: mismatched + wrong drafts (ours)
hugruby/mathstral-7b-grpo-no-draft

Text Generation • Updated Jun 11 • 6

Note Baseline: no-draft GRPO
hugruby/mathstral-7b-matched-wrong-drafts

Text Generation • Updated Jun 11 • 6

Note Ablation: matched + wrong
hugruby/mathstral-7b-mismatched-correct-drafts

Text Generation • Updated 18 days ago • 16

Note Ablation: mismatched + correct
hugruby/mismatched-wrong-drafts

Viewer • Updated 18 days ago • 35.6k • 165

Note 8,888 MATH problems; 4 draft conditions + drafts.json