anaghaa/resonance
Viewer • Updated • 577k • 88
Resonance
Models that predict reaction products by reasoning: Qwen3-8B and Qwen3.8-27B distilled from Kimi K3, then self-trained. With all data.
Note All data: teacher traces, training sets, test outputs, probe samples
anaghaa/Resonance-8B-S1
Text Generation • 8B • Updated • 5Note 8B after teacher SFT (31.0%)
anaghaa/Resonance-8B-S2
Text Generation • 8B • Updated • 6Note 8B after rejection-sampling SFT (38.9%)
anaghaa/Resonance-8B-S3
Text Generation • 8B • Updated • 5Note Final 8B model (42.8% USPTO-50k top-1)
anaghaa/Resonance-8B-skip-S2
Text Generation • 8B • Updated • 6Note Ablation: DPO directly on 8B S1 (37.1%)
anaghaa/Resonance-27B-S1
Image-Text-to-Text • 3.05M • Updated • 17Note 27B after teacher SFT (79.0%)
anaghaa/Resonance-27B-S2
Image-Text-to-Text • 27B • Updated • 17Note 27B after rejection-sampling SFT (79.3%)
anaghaa/Resonance-27B-S3
Image-Text-to-Text • 27B • Updated • 16Note Final 27B model (79.6% USPTO-50k top-1)
anaghaa/Resonance-27B-skip-S2
Image-Text-to-Text • 27B • Updated • 16Note Ablation: DPO directly on 27B S1 (78.8%)