Filter higher quality pretraining data.

Created on April 15|Last edited on April 15
Comment
﻿
Instead of pretraining on all the observation data we had, I restricted to the smaller (~1K) trials of Will's data that did have >0.5 R2. This did not help downstream performance.
﻿
﻿
eval_kinematic_r2
eval_kinematic_r2
123456789102030405060708090100epoch00.10.20.30.40.5
human_nofilter-frag-observation_CRS02bLab_session_1914.*-4603o4m0   40
human_nofilter-frag-observation_CRS02bLab_session_1922.*-ccyaz4eb   40
human_nofilter-frag-observation_CRS02bLab_session_1913.*-g1rzztss   40
human_nofilter-frag-observation_CRS02bLab_session_1908.*-lnyk5yfy   40
human_nofilter-frag-observation_CRS02bLab_session_1925.*-y9mlddkn   40
human_nofilter-frag-observation_CRS02bLab_session_1926.*-lglsgrgs   40
human_nofilter-frag-observation_CRS02bLab_session_1904.*-0q4lowjd   40
human_nofilter-frag-observation_CRS02bLab_session_1903.*-mtf1c299   40
human-frag-observation_CRS02bLab_session_1922.*-6wzarpb3   40
human-frag-observation_CRS02bLab_session_1926.*-av4qoxs0   40
human-frag-observation_CRS02bLab_session_1925.*-kwh4fcfs   40
human-frag-observation_CRS02bLab_session_1913.*-hosr6iwh   40
human-frag-observation_CRS02bLab_session_1914.*-lno89dux   40
human-frag-observation_CRS02bLab_session_1908.*-t5fahopi   40
human-frag-observation_CRS02bLab_session_1904.*-326ka9aa   40
human-frag-observation_CRS02bLab_session_1903.*-6i64uc4d   40
Run set16
﻿
﻿
Add a comment
Filter higher quality pretraining data.

﻿
