[17:29:35] <sabya>	 o/ halfak 
[17:49:47] <halfak>	 o/ sabya 
[17:50:12] <sabya>	 did you get a chance to look at the results?
[17:51:21] <halfak>	 I quickly scanned your notes, but not in depth yet. 
[17:52:13] <sabya>	 ok. in short, as we add HV features, ROC AUC decreases.
[17:52:43] <halfak>	 sabya, we might need to tune for other parameterizations of the GB classifier
[17:54:22] <sabya>	 ok
[17:54:37] * halfak reads more carefully
[18:03:04] <halfak>	 sabya, looks like the model w/o hash features isn't doing as good as our current model without hashed features. 
[18:03:33] <halfak>	 I didn't think we made any read improvements since then. 
[18:03:39] <halfak>	 So I'd like to try tuning.
[18:05:14] <halfak>	 Ahh yes.  It looks like you are parameterizing the GB model differently than we do for production. 
[18:05:49] <halfak>	 See https://ores.wikimedia.org/v2/scores/enwiki/damaging/?model_info=params
[18:06:09] * sabya reads the link
[18:06:21] <halfak>	 n_estimators = 700, max_depth = 7, learning_rate = 0.01
[18:07:33] <halfak>	 https://github.com/wiki-ai/editquality/blob/master/tuning_reports/enwiki.damaging.md
[18:07:51] <halfak>	 This is a tuning report from a recent version of the model.
[18:09:09] <sabya>	 ok
[18:14:11] <sabya>	 gotta go now. will try with the above tuning params tomorrow. bye halfak 
[18:14:22] <halfak>	 o/ sabya 
[18:14:25] <halfak>	 Have a good one
[18:25:00] * halfak looks around for YuviFanta
[18:25:15] <halfak>	 (orange preferably) 
[18:26:53] <hare>	 YuviFanta??
[18:27:05] <hare>	 I like the pineapple Fanta. Yes, it actually tastes good!