Add rejection sampling analysis

[vwxyzjn]

🚀 Dataset loaded, starting to analyze...
chosen:    mean length = 163.50656107997827
rejected:  mean length = 150.02915225944915
reference: mean length = 811.4574606202721
key='allenai/llama-3-tulu-2-8b-uf-mean-rm'
chosen:    mean score = 26.952228895910338
           std score = 3.0159979740827905
rejected:  mean score = 23.47721361929176
           std score = 3.226889784341013
reference: mean score = 30.24208389637402
           std score = 4.438492691848654
reference is better than chosen

So I realized the issue is that we generate only 53 tokens by default. We should change it to a much more sensible number, prob 2048 @nouhadziri

open-instruct/open_instruct/generation.py

Line 43 in 1515330

response_length: int = 53

Let's merge this after your PR #252 and I can add documentation

[nouhadziri]

This explains the performance!! I haven't had a chance to play with different generation parameters yet but increasing the length will definitely boost numbers.

[nouhadziri]

LGTM