--- tags: - c-vis - ablation - grpo - qwen3.5 library_name: transformers --- # C-VIS ablation checkpoints (cvis_ablate) Solver weights from the C-VIS ablation study (branch `cvis_ablate`). Each folder holds one finished round: a full Qwen3.5-9B fine-tune after 160 GRPO optimizer steps, trained with ms-swift. Layout: `/round_/` contains the safetensors shards, config, tokenizer, processor, chat template and `trainer_state.json` (training log history). Optimizer states are not included. | Experiment | Round | MathVision | MathVista | MMMU | Folder | | --- | --- | --- | --- | --- | --- | | a4_fixed_images | 1 | 50.66% | 81.80% | 68.67% | `a4_fixed_images/round_1` | | a4_fixed_images | 2 | 51.64% | 82.60% | 68.44% | `a4_fixed_images/round_2` | | a4_fixed_images | 3 | 53.95% | 82.80% | 68.22% | `a4_fixed_images/round_3` | | a4_fixed_images | 4 | 52.30% | 82.90% | 68.56% | `a4_fixed_images/round_4` | | c1_no_history | 1 | 53.29% | 82.60% | 69.22% | `c1_no_history/round_1` | | c1_no_history | 2 | 53.29% | 82.40% | 69.00% | `c1_no_history/round_2` | | c1_no_history | 3 | 54.28% | 83.20% | 70.11% | `c1_no_history/round_3` | | c1_no_history | 4 | 50.33% | 82.80% | 70.89% | `c1_no_history/round_4` | | c2_uniform_difficulty | 1 | 52.63% | 81.80% | 70.33% | `c2_uniform_difficulty/round_1` | | c2_uniform_difficulty | 2 | 54.61% | 83.20% | 71.44% | `c2_uniform_difficulty/round_2` | | c2_uniform_difficulty | 3 | 54.61% | 82.90% | 69.00% | `c2_uniform_difficulty/round_3` | | c2_uniform_difficulty | 4 | 51.97% | 82.50% | 69.78% | `c2_uniform_difficulty/round_4` | | c3_uniform_category | 1 | 53.29% | 83.70% | 70.11% | `c3_uniform_category/round_1` | | c3_uniform_category | 2 | 53.95% | 81.80% | 70.67% | `c3_uniform_category/round_2` | | c3_uniform_category | 3 | 53.95% | 82.80% | 69.78% | `c3_uniform_category/round_3` | | c3_uniform_category | 4 | 54.61% | 83.40% | 71.44% | `c3_uniform_category/round_4` | | c4_no_difficulty_filter | 1 | 54.93% | 82.00% | 69.78% | `c4_no_difficulty_filter/round_1` | | c4_no_difficulty_filter | 2 | 56.25% | 82.10% | 69.00% | `c4_no_difficulty_filter/round_2` | | c4_no_difficulty_filter | 3 | 55.59% | 83.20% | 70.00% | `c4_no_difficulty_filter/round_3` | | c4_no_difficulty_filter | 4 | 57.57% | 82.30% | 70.11% | `c4_no_difficulty_filter/round_4` | Experiments: - `c1_no_history`: curriculum without history. - `c2_uniform_difficulty`: uniform difficulty curriculum. - `c3_uniform_category`: uniform category curriculum. - `c4_no_difficulty_filter`: no difficulty filter. - `a4_fixed_images`: fixed image pool. Scores are the full benchmark splits used by the C-VIS evaluation (MathVision 304, MathVista 1000, MMMU 900), sampled at temperature 0.7. More rounds are added as runs finish. Load one round with `from_pretrained("", subfolder="c4_no_difficulty_filter/round_4")`.