83 lines
3.2 KiB
YAML
83 lines
3.2 KiB
YAML
|
model_name: Tucano2-qwen-3.7B-Base
|
||
|
|
results:
|
||
|
|
niah_pt_multikey_1_1024: 0.596
|
||
|
|
niah_pt_multikey_1_1024_stderr: 0.021966635293832883
|
||
|
|
niah_pt_multikey_1_2048: 0.614
|
||
|
|
niah_pt_multikey_1_2048_stderr: 0.021793529219281196
|
||
|
|
niah_pt_multikey_1_4096: 0.436
|
||
|
|
niah_pt_multikey_1_4096_stderr: N/A
|
||
|
|
niah_pt_multikey_1_alias: " - niah_pt_multikey_1"
|
||
|
|
niah_pt_multikey_2_1024: 0.67
|
||
|
|
niah_pt_multikey_2_1024_stderr: 0.021049612166134782
|
||
|
|
niah_pt_multikey_2_2048: 0.682
|
||
|
|
niah_pt_multikey_2_2048_stderr: 0.020847571620814014
|
||
|
|
niah_pt_multikey_2_4096: 0.608
|
||
|
|
niah_pt_multikey_2_4096_stderr: N/A
|
||
|
|
niah_pt_multikey_2_alias: " - niah_pt_multikey_2"
|
||
|
|
niah_pt_multikey_3_1024: 0.842
|
||
|
|
niah_pt_multikey_3_1024_stderr: 0.0163280498045799
|
||
|
|
niah_pt_multikey_3_2048: 0.856
|
||
|
|
niah_pt_multikey_3_2048_stderr: 0.015716934945725784
|
||
|
|
niah_pt_multikey_3_4096: 0.786
|
||
|
|
niah_pt_multikey_3_4096_stderr: N/A
|
||
|
|
niah_pt_multikey_3_alias: " - niah_pt_multikey_3"
|
||
|
|
niah_pt_multiquery_1024: 0.5815
|
||
|
|
niah_pt_multiquery_1024_stderr: 0.015245666949052864
|
||
|
|
niah_pt_multiquery_2048: 0.5405
|
||
|
|
niah_pt_multiquery_2048_stderr: 0.014383437229433934
|
||
|
|
niah_pt_multiquery_4096: 0.493
|
||
|
|
niah_pt_multiquery_4096_stderr: N/A
|
||
|
|
niah_pt_multiquery_alias: " - niah_pt_multiquery"
|
||
|
|
niah_pt_multivalue_1024: 0.5685
|
||
|
|
niah_pt_multivalue_1024_stderr: 0.015209475872383808
|
||
|
|
niah_pt_multivalue_2048: 0.3425
|
||
|
|
niah_pt_multivalue_2048_stderr: 0.014868747738751817
|
||
|
|
niah_pt_multivalue_4096: 0.355
|
||
|
|
niah_pt_multivalue_4096_stderr: N/A
|
||
|
|
niah_pt_multivalue_alias: " - niah_pt_multivalue"
|
||
|
|
niah_pt_single_1_1024: 0.63
|
||
|
|
niah_pt_single_1_1024_stderr: 0.021613289165165816
|
||
|
|
niah_pt_single_1_2048: 0.578
|
||
|
|
niah_pt_single_1_2048_stderr: 0.022109039310618563
|
||
|
|
niah_pt_single_1_4096: 0.462
|
||
|
|
niah_pt_single_1_4096_stderr: N/A
|
||
|
|
niah_pt_single_1_alias: " - niah_pt_single_1"
|
||
|
|
niah_pt_single_2_1024: 0.538
|
||
|
|
niah_pt_single_2_1024_stderr: 0.022318338119870523
|
||
|
|
niah_pt_single_2_2048: 0.548
|
||
|
|
niah_pt_single_2_2048_stderr: 0.02227969410784354
|
||
|
|
niah_pt_single_2_4096: 0.566
|
||
|
|
niah_pt_single_2_4096_stderr: N/A
|
||
|
|
niah_pt_single_2_alias: " - niah_pt_single_2"
|
||
|
|
niah_pt_single_3_1024: 0.716
|
||
|
|
niah_pt_single_3_1024_stderr: 0.020186703693570777
|
||
|
|
niah_pt_single_3_2048: 0.726
|
||
|
|
niah_pt_single_3_2048_stderr: 0.019966103540279518
|
||
|
|
niah_pt_single_3_4096: 0.702
|
||
|
|
niah_pt_single_3_4096_stderr: N/A
|
||
|
|
niah_pt_single_3_alias: " - niah_pt_single_3"
|
||
|
|
ruler_pt_4096: 0.5871696969696969
|
||
|
|
ruler_pt_4096_stderr: N/A
|
||
|
|
ruler_pt_alias: ruler_pt
|
||
|
|
ruler_pt_cwe_1024: 0.3028
|
||
|
|
ruler_pt_cwe_1024_stderr: 0.0064114376754463
|
||
|
|
ruler_pt_cwe_2048: 0.253
|
||
|
|
ruler_pt_cwe_2048_stderr: 0.006635515060659607
|
||
|
|
ruler_pt_cwe_4096: 0.3358
|
||
|
|
ruler_pt_cwe_4096_stderr: N/A
|
||
|
|
ruler_pt_cwe_alias: " - ruler_pt_cwe"
|
||
|
|
ruler_pt_fwe_1024: 0.864
|
||
|
|
ruler_pt_fwe_1024_stderr: 0.010004852975181538
|
||
|
|
ruler_pt_fwe_2048: 0.8266666666666667
|
||
|
|
ruler_pt_fwe_2048_stderr: 0.011102646519713471
|
||
|
|
ruler_pt_fwe_4096: 0.7346666666666666
|
||
|
|
ruler_pt_fwe_4096_stderr: N/A
|
||
|
|
ruler_pt_fwe_alias: " - ruler_pt_fwe"
|
||
|
|
ruler_pt_vt_1024: 0.9816
|
||
|
|
ruler_pt_vt_1024_stderr: 0.004767424532527671
|
||
|
|
ruler_pt_vt_2048: 0.9708
|
||
|
|
ruler_pt_vt_2048_stderr: 0.006233171331649675
|
||
|
|
ruler_pt_vt_4096: 0.9803999999999999
|
||
|
|
ruler_pt_vt_4096_stderr: N/A
|
||
|
|
ruler_pt_vt_alias: " - ruler_pt_vt"
|