| license: apache-2.0 | |
| base_model: Qwen/Qwen3-VL-4B-Instruct | |
| tags: | |
| - robometer | |
| - reward-model | |
| - rbm | |
| - armnet | |
| - molmoact | |
| library_name: transformers | |
| # Robometer-4B FFT finetuned on both so101 datasets (Armnet so101 + MolmoACT2 so101) | |
| Full fine-tune (FFT, no LoRA) of [Robometer-4B](https://huggingface.co/robometer/Robometer-4B) on **both** SO-101 datasets: | |
| - Armnet benchmark so101 (`villekuosmanen_armnetbench_robometer_v01_so101`) | |
| - MolmoACT2 so101 (`ykorkmaz_molmoact2_so100_101_rbm_molmoact2_so100_101`) | |
| Qwen3-VL-4B backbone, 1500 steps on 4x H200. This is the best checkpoint (step 750). | |
| ## Key result: training on both so101 datasets improves generalization | |
| | Metric | armnet-only finetune | this (both so101) | | |
| |---|---|---| | |
| | Armnet so101 reward-alignment Pearson | 0.766 | 0.782 | | |
| | Armnet so101 policy-ranking Kendall | 0.973 | 0.94 | | |
| | Molmoact so101 reward-alignment Pearson (held-out) | 0.751 | **0.902** | | |
| Training on both datasets raised molmoact so101 Pearson from 0.75 → 0.90 while keeping armnet performance roughly intact. | |