TRAIN_PR_REWRITE_CLASSIFICATION.md

Train Pure_Reasoning — rewrite classification

Per-task two-phase scan: phase 1 inspects up to 50 samples to detect which buckets the task has any samples in; phase 2 (only triggered if a bucket was seen but not yet filled to 2) scans up to 500 samples to find rare paraphrases. A sample is rewritten if its prompt differs from prompt_original; unchanged otherwise (covers both already-clean tasks with no prompt_original and rewrite-engine no-ops). The two new HF splits (train_pr_rewritten, train_pr_unchanged) carry these samples.

Summary: 51 train PR tasks total. Rewritten-only: 28. Unchanged-only: 20. Both buckets present: 3. Neither (no samples scanned): 0.

Note on the 'both buckets' tasks: O-10, O-16, O-17 contain multiple distinct prompt-paraphrase templates per task; some templates trip a leak-rewrite rule and others do not. Verified by full 10,000-sample scan: O-10 is 97.8% rewritten / 2.2% unchanged (4 paraphrase templates), O-16 is 49.8% / 50.2% (4 templates), O-17 is 24.7% / 75.3% (4 templates). All three contribute 2 samples to each split.

Tasks contributing to BOTH splits (3 tasks)

Task	Scanned	rewritten seen	unchanged seen	added → rewritten	added → unchanged
O-10_shape_outline_fill_data-generator	75	73	2	2	2
O-16_color_addition_data-generator	5	3	2	2	2
O-17_color_subtraction_data-generator	4	2	2	2	2

Tasks contributing ONLY to train_pr_rewritten (28 tasks)

Task	Scanned	rewritten seen	unchanged seen	added → rewritten	added → unchanged
G-131_select_next_figure_increasing_size_sequence_data-generator	50	50	0	2	0
G-133_select_next_figure_decreasing_size_sequence_data-generator	50	50	0	2	0
G-134_select_next_figure_large_small_alternating_sequence_data-generator	50	50	0	2	0
G-21B_multiple_occlusions_vertical_behind_data-generator	50	50	0	2	0
G-21_multiple_occlusions_vertical_data-generator	50	50	0	2	0
G-35_hit_target_after_bounce_data-generator	50	50	0	2	0
G-36B_multiple_occlusions_horizontal_behind_data-generator	50	50	0	2	0
G-36_multiple_occlusions_horizontal_data-generator	50	50	0	2	0
G-41_grid_highest_cost_data-generator	50	50	0	2	0
G-44_bfs_data-generator	50	50	0	2	0
G-45_key_door_matching_data-generator	50	50	0	2	0
G-46_find_keys_and_open_doors_data-generator	50	50	0	2	0
G-48_multiple_bounces_data-generator	50	50	0	2	0
O-12_shape_color_then_scale_data-generator	50	50	0	2	0
O-13_shape_outline_then_move_data-generator	50	50	0	2	0
O-14_shape_scale_then_outline_data-generator	50	50	0	2	0
O-15_ball_bounces_given_time_data-generator	50	50	0	2	0
O-18_glass_refraction_data-generator	50	50	0	2	0
O-1_color_mixing_data-generator	50	50	0	2	0
O-21_construction_blueprint_data-generator	50	50	0	2	0
O-23_domino_chain_branch_path_prediction_data-generator	50	50	0	2	0
O-24_domino_chain_gap_analysis_data-generator	50	50	0	2	0
O-29_ballcolor_data-generator	50	50	0	2	0
O-31_ball_eating_data-generator	50	50	0	2	0
O-45_sequence_completion_data-generator	50	50	0	2	0
O-75_communicating_vessels_data-generator	50	50	0	2	0
O-7_shape_color_change_data-generator	50	50	0	2	0
O-87_fluid_diffusion_reasoning_data-generator	50	50	0	2	0

Tasks contributing ONLY to train_pr_unchanged (20 tasks)

Task	Scanned	rewritten seen	unchanged seen	added → rewritten	added → unchanged
G-12_grid_obtaining_award_data-generator	50	0	50	0	2
G-132_find_fragment_for_gap_filling_data-generator	50	0	50	0	2
G-13_grid_number_sequence_data-generator	50	0	50	0	2
G-14_grid_color_sequence_data-generator	50	0	50	0	2
G-15_grid_avoid_obstacles_data-generator	50	0	50	0	2
G-16_grid_go_through_block_data-generator	50	0	50	0	2
G-17_grid_avoid_red_block_data-generator	50	0	50	0	2
G-18_grid_shortest_path_data-generator	50	0	50	0	2
G-31_directed_graph_navigation_data-generator	50	0	50	0	2
G-32_undirected_graph_navigation_data-generator	50	0	50	0	2
G-43_understand_scene_structure_data-generator	50	0	50	0	2
G-51_predict_next_color_data-generator	50	0	50	0	2
O-18B_glass_refraction_inverse_data-generator	50	0	50	0	2
O-19B_mirror_reflection_inverse_data-generator	50	0	50	0	2
O-19_mirror_reflection_data-generator	50	0	50	0	2
O-44_rotation_puzzle_data-generator	50	0	50	0	2
O-47_sliding_puzzle_data-generator	50	0	50	0	2
O-52_traffic_light_data-generator	50	0	50	0	2
O-53_clock_data-generator	50	0	50	0	2
O-8_shape_rotation_data-generator	50	0	50	0	2

---

Generated by scripts/vbvr_reorg/build_train_pr_diff_splits.py. Total scan-cap: 50 samples/task; per-bucket pick: 2.