experiments / counter_consistent_results_before_ai2thor_fixed.txt
ch-min's picture
Add files using upload-large-folder tool
3404d44 verified
======================================================================
Model: molmo-7B-O-0924
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 328 487 67.4%
FAR counter 16 63 25.4%
FAR ambiguous 22 44 50.0%
CLOSE consistent 298 500 59.6%
CLOSE counter 30 69 43.5%
CLOSE ambiguous 17 43 39.5%
----------------------------------------------------
TOTAL consistent 626 987 63.4%
TOTAL counter 46 132 34.8%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 28.6%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 86 / 132
======================================================================
Model: molmo-7B-O-0924-data_scale_exp_80k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 271 487 55.6%
FAR counter 19 63 30.2%
FAR ambiguous 19 44 43.2%
CLOSE consistent 324 500 64.8%
CLOSE counter 20 69 29.0%
CLOSE ambiguous 13 43 30.2%
----------------------------------------------------
TOTAL consistent 595 987 60.3%
TOTAL counter 39 132 29.5%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 30.7%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 93 / 132
======================================================================
Model: molmo-7B-O-0924-data_scale_exp_400k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 317 487 65.1%
FAR counter 15 63 23.8%
FAR ambiguous 17 44 38.6%
CLOSE consistent 300 500 60.0%
CLOSE counter 21 69 30.4%
CLOSE ambiguous 12 43 27.9%
----------------------------------------------------
TOTAL consistent 617 987 62.5%
TOTAL counter 36 132 27.3%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 35.2%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 96 / 132
======================================================================
Model: molmo-7B-O-0924-data_scale_exp_800k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 320 487 65.7%
FAR counter 23 63 36.5%
FAR ambiguous 17 44 38.6%
CLOSE consistent 320 500 64.0%
CLOSE counter 23 69 33.3%
CLOSE ambiguous 19 43 44.2%
----------------------------------------------------
TOTAL consistent 640 987 64.8%
TOTAL counter 46 132 34.8%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 30.0%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 86 / 132
======================================================================
Model: molmo-7B-O-0924-data_scale_exp_2m
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 313 487 64.3%
FAR counter 19 63 30.2%
FAR ambiguous 18 44 40.9%
CLOSE consistent 328 500 65.6%
CLOSE counter 34 69 49.3%
CLOSE ambiguous 20 43 46.5%
----------------------------------------------------
TOTAL consistent 641 987 64.9%
TOTAL counter 53 132 40.2%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 24.8%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 79 / 132
======================================================================
Model: NVILA-Lite-2B
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 73 487 15.0%
FAR counter 3 63 4.8%
FAR ambiguous 5 44 11.4%
CLOSE consistent 77 500 15.4%
CLOSE counter 8 69 11.6%
CLOSE ambiguous 7 43 16.3%
----------------------------------------------------
TOTAL consistent 150 987 15.2%
TOTAL counter 11 132 8.3%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 6.9%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 121 / 132
======================================================================
Model: NVILA-Lite-2B-data-scale-exp-80k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 317 487 65.1%
FAR counter 2 63 3.2%
FAR ambiguous 11 44 25.0%
CLOSE consistent 248 500 49.6%
CLOSE counter 19 69 27.5%
CLOSE ambiguous 12 43 27.9%
----------------------------------------------------
TOTAL consistent 565 987 57.2%
TOTAL counter 21 132 15.9%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 41.3%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 111 / 132
======================================================================
Model: NVILA-Lite-2B-data-scale-exp-400k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 319 487 65.5%
FAR counter 12 63 19.0%
FAR ambiguous 18 44 40.9%
CLOSE consistent 280 500 56.0%
CLOSE counter 33 69 47.8%
CLOSE ambiguous 16 43 37.2%
----------------------------------------------------
TOTAL consistent 599 987 60.7%
TOTAL counter 45 132 34.1%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 26.6%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 87 / 132
======================================================================
Model: NVILA-Lite-2B-data-scale-exp-800k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 336 487 69.0%
FAR counter 17 63 27.0%
FAR ambiguous 19 44 43.2%
CLOSE consistent 283 500 56.6%
CLOSE counter 34 69 49.3%
CLOSE ambiguous 13 43 30.2%
----------------------------------------------------
TOTAL consistent 619 987 62.7%
TOTAL counter 51 132 38.6%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 24.1%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 81 / 132
======================================================================
Model: NVILA-Lite-2B-data-scale-exp-2m
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 324 487 66.5%
FAR counter 19 63 30.2%
FAR ambiguous 21 44 47.7%
CLOSE consistent 271 500 54.2%
CLOSE counter 35 69 50.7%
CLOSE ambiguous 16 43 37.2%
----------------------------------------------------
TOTAL consistent 595 987 60.3%
TOTAL counter 54 132 40.9%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 19.4%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 78 / 132
======================================================================
Model: RoboRefer-2B-SFT
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 425 487 87.3%
FAR counter 30 63 47.6%
FAR ambiguous 29 44 65.9%
CLOSE consistent 432 500 86.4%
CLOSE counter 49 69 71.0%
CLOSE ambiguous 32 43 74.4%
----------------------------------------------------
TOTAL consistent 857 987 86.8%
TOTAL counter 79 132 59.8%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 27.0%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 53 / 132
======================================================================
Model: Qwen2.5-VL-3B-Instruct
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 274 487 56.3%
FAR counter 15 63 23.8%
FAR ambiguous 16 44 36.4%
CLOSE consistent 263 500 52.6%
CLOSE counter 28 69 40.6%
CLOSE ambiguous 13 43 30.2%
----------------------------------------------------
TOTAL consistent 537 987 54.4%
TOTAL counter 43 132 32.6%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 21.8%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 89 / 132
======================================================================
Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_80k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 268 487 55.0%
FAR counter 15 63 23.8%
FAR ambiguous 12 44 27.3%
CLOSE consistent 231 500 46.2%
CLOSE counter 25 69 36.2%
CLOSE ambiguous 13 43 30.2%
----------------------------------------------------
TOTAL consistent 499 987 50.6%
TOTAL counter 40 132 30.3%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 20.3%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 92 / 132
======================================================================
Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_400k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 276 487 56.7%
FAR counter 13 63 20.6%
FAR ambiguous 16 44 36.4%
CLOSE consistent 241 500 48.2%
CLOSE counter 23 69 33.3%
CLOSE ambiguous 13 43 30.2%
----------------------------------------------------
TOTAL consistent 517 987 52.4%
TOTAL counter 36 132 27.3%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 25.1%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 96 / 132
======================================================================
Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_800k
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 298 487 61.2%
FAR counter 12 63 19.0%
FAR ambiguous 15 44 34.1%
CLOSE consistent 251 500 50.2%
CLOSE counter 23 69 33.3%
CLOSE ambiguous 11 43 25.6%
----------------------------------------------------
TOTAL consistent 549 987 55.6%
TOTAL counter 35 132 26.5%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 29.1%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 97 / 132
======================================================================
Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_2m
======================================================================
Category Type Correct Total Accuracy
----------------------------------------------------
FAR consistent 313 487 64.3%
FAR counter 6 63 9.5%
FAR ambiguous 14 44 31.8%
CLOSE consistent 284 500 56.8%
CLOSE counter 26 69 37.7%
CLOSE ambiguous 15 43 34.9%
----------------------------------------------------
TOTAL consistent 597 987 60.5%
TOTAL counter 32 132 24.2%
๐Ÿ“Š Accuracy Gap (Consistent - Counter): 36.2%p
โ†’ Gap์ด ํด์ˆ˜๋ก ๋ชจ๋ธ์ด 2D heuristic์— ์˜์กดํ•˜๋Š” ๊ฒฝํ–ฅ
๐Ÿ” Counter Examples ์˜ค๋‹ต ์ˆ˜: 100 / 132
==============================================================================
MODEL COMPARISON
==============================================================================
Model Consistent Counter Gap
------------------------------------------------------------------------------
molmo-7B-O-0924 63.4% 34.8% +28.6%p
molmo-7B-O-0924-data_scale_exp_80k 60.3% 29.5% +30.7%p
molmo-7B-O-0924-data_scale_exp_400k 62.5% 27.3% +35.2%p
molmo-7B-O-0924-data_scale_exp_800k 64.8% 34.8% +30.0%p
molmo-7B-O-0924-data_scale_exp_2m 64.9% 40.2% +24.8%p
NVILA-Lite-2B 15.2% 8.3% +6.9%p
NVILA-Lite-2B-data-scale-exp-80k 57.2% 15.9% +41.3%p
NVILA-Lite-2B-data-scale-exp-400k 60.7% 34.1% +26.6%p
NVILA-Lite-2B-data-scale-exp-800k 62.7% 38.6% +24.1%p
NVILA-Lite-2B-data-scale-exp-2m 60.3% 40.9% +19.4%p
RoboRefer-2B-SFT 86.8% 59.8% +27.0%p
Qwen2.5-VL-3B-Instruct 54.4% 32.6% +21.8%p
Qwen2.5-VL-3B-Instruct-data_scale_exp_80k 50.6% 30.3% +20.3%p
Qwen2.5-VL-3B-Instruct-data_scale_exp_400k 52.4% 27.3% +25.1%p
Qwen2.5-VL-3B-Instruct-data_scale_exp_800k 55.6% 26.5% +29.1%p
Qwen2.5-VL-3B-Instruct-data_scale_exp_2m 60.5% 24.2% +36.2%p