svgicons_curriculum_v2_minimal

# SVG Icons Curriculum SFT Dataset v2 (Minimal) Minimal subset of [svgicons_curriculum_v2](../svgicons_curriculum_v2/), containing only **new_prompt** experiments at **10%, 50%, and 100%** data scales with all three training strategies. ## What's included - **Prompt variant**: `new_prompt` only (pure visual complexity scoring, no path-count anchors) - **Scales**: `10pct` (7,836), `50pct` (39,180), `full` (78,361) - **Strategies**: `baseline`, `stage1`, `stage2` (9 training configs total) - **Images**: 153,309 PNGs (from 805,079 in full version, ~19%) - **Eval/Test**: `val.json` (6,094), `test.json` (2,594) -- shared, unchanged - **Scores**: `scores/new_prompt/` raw VLM difficulty scores - **Examples**: LlamaFactory training YAML templates ## Directory Structure ``` svgicons_curriculum_v2_minimal/ ├── images/ # 153,309 PNGs (320x320), 需解压 images.tar.gz ├── new_prompt/ # 含 decomposition 的 curriculum（原始方案） │ ├── {10pct,50pct,full}_{baseline,stage1,stage2}_train.json │ └── pool_summary.json ├── new_prompt_aligned/ # 含 decomposition 的 curriculum（stage1 数量对齐 baseline） │ ├── {10pct,50pct,full}_stage1_train.json │ └── {10pct,50pct,full}_stage2_train.json ├── nodec_path/ # 无分解 curriculum：按 path 数排序 │ ├── full_stage1_train.json # 23,508 条（n_paths 少 → 多，前 30%） │ └── full_stage2_train.json # 78,361 条（全部） ├── nodec_score/ # 无分解 curriculum：按 difficulty score 排序 │ ├── full_stage1_train.json # 23,508 条（score 低 → 高，前 30%） │ └── full_stage2_train.json # 78,361 条（全部） ├── scores/new_prompt/ │ ├── train_scores.jsonl │ ├── val_scores.jsonl │ └── test_scores.jsonl ├── examples/ │ ├── qwen3vl_8b_stage1_lora.yaml │ └── qwen3vl_8b_stage2_lora.yaml ├── val.json ├── test.json ├── debug_train.json ├── debug_test.json ├── dataset_info.json # 19 LlamaFactory dataset entries └── README.md ``` ## Training Sizes ### 含 decomposition 的 curriculum（new_prompt/） | Scale | Baseline | Stage1 | Stage2 | |---|---:|---:|---:| | 10pct | 7,836 | 4,300 | 14,335 | | 50pct | 39,180 | 21,609 | 72,033 | | full | 78,361 | 43,386 | 144,621 | ### 含 decomposition 的 curriculum，stage1 对齐 baseline（new_prompt_aligned/） Stage1 从 pool 中按 eff_diff 升序取与 baseline 等量的样本，方便直接对比。 | Scale | Stage1 (= Baseline) | Stage2 | |---|---:|---:| | 10pct | 7,836 | 14,335 | | 50pct | 39,180 | 72,033 | | full | 78,361 | 144,621 | ### 无分解 curriculum（nodec_path/ 和 nodec_score/） 只使用原始完整样本（无 prefix decomposition），curriculum 仅在训练阶段层面体现从易到难。 | Curriculum | Stage1 (前 30%) | Stage2 (全部) | 排序依据 | |---|---:|---:|---| | nodec_path | 23,508 | 78,361 | n_paths 升序（组内 seed=42 随机） | | nodec_score | 23,508 | 78,361 | difficulty score 升序（组内 seed=42 随机） | **nodec_path stage1 分布**: {1: 2030, 2: 5070, 3: 7205, 4: 7829, 5: 1374} **nodec_score stage1 分布**: {score1: 934, score2: 3411, score3: 19163} ### LlamaFactory Dataset Names | Dataset Name | File | |---|---| | `svgicons_nodec_path_full_stage1` | `nodec_path/full_stage1_train.json` | | `svgicons_nodec_path_full_stage2` | `nodec_path/full_stage2_train.json` | | `svgicons_nodec_score_full_stage1` | `nodec_score/full_stage1_train.json` | | `svgicons_nodec_score_full_stage2` | `nodec_score/full_stage2_train.json` | | `svgicons_new_prompt_aligned_10pct_stage1` | `new_prompt_aligned/10pct_stage1_train.json` | | `svgicons_new_prompt_aligned_10pct_stage2` | `new_prompt_aligned/10pct_stage2_train.json` | | `svgicons_new_prompt_aligned_50pct_stage1` | `new_prompt_aligned/50pct_stage1_train.json` | | `svgicons_new_prompt_aligned_50pct_stage2` | `new_prompt_aligned/50pct_stage2_train.json` | | `svgicons_new_prompt_aligned_full_stage1` | `new_prompt_aligned/full_stage1_train.json` | | `svgicons_new_prompt_aligned_full_stage2` | `new_prompt_aligned/full_stage2_train.json` | ## Usage Copy `dataset_info.json` to your LlamaFactory data directory and point to this folder. ```bash cp dataset_info.json /path/to/LlamaFactory/data/ # 无分解 path curriculum CUDA_VISIBLE_DEVICES=0,1 FORCE_TORCHRUN=1 llamafactory-cli train train/qwen3vl_8b_path_stage1_lora.yaml CUDA_VISIBLE_DEVICES=0,1 FORCE_TORCHRUN=1 llamafactory-cli train train/qwen3vl_8b_path_stage2_lora.yaml # 无分解 score curriculum CUDA_VISIBLE_DEVICES=0,1 FORCE_TORCHRUN=1 llamafactory-cli train train/qwen3vl_8b_score_stage1_lora.yaml CUDA_VISIBLE_DEVICES=0,1 FORCE_TORCHRUN=1 llamafactory-cli train train/qwen3vl_8b_score_stage2_lora.yaml ``` ## Source Derived from `svgicons_curriculum_v2` -- see its README for full documentation on curriculum design, augmentation, and scoring methodology.