File size: 2,855 Bytes
cafcb62 446ec5e debae52 cafcb62 403b3aa 8c5f474 60579f7 cafcb62 71c5c9b cafcb62 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 | ---
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: `<experiment>/round_<N>/` 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("<repo>", subfolder="c4_no_difficulty_filter/round_4")`.
|