Leon

L/S

[Leon @LeonLondon9]

@AscendedYield Romeoo

[Leon @LeonLondon9]

@kalomaze What's the intuition behind the binary classification ^4 scaling? If model gets 4/8 correct, it's guessing, don't want to reward this behaviour, if it's 7/8, it's picking up smth?

[Leon @LeonLondon9]

@kalomaze Which hyperparams? I keep getting junk text after 100 steps (500 samples) and I can't tell if it's a sampling issue. Lr 5e-7. Happens even before the Lr decay. generally happens when the format rewards saturate

[Leon @LeonLondon9]

@nrehiew_ Liger kernel, 8 bit Adam, using mini batch logit processing from Andy pr in trl library. Your best bet imo

[Leon @LeonLondon9]

@alkimiadev @UnslothAI What kind of rewards are you using for text classification? Straight up 1/0 reward? Any intermediate reward steps?

Thanks for the info, it's very interesting.

[Leon @LeonLondon9]

@michael_chomsky @corbtt @simonw Why is it crazy?

[Leon @LeonLondon9]

@W98AB @tango_contango @quantymacro Still fair few people trading Europe working today and took 1st may off instead

[Leon @LeonLondon9]

@goodalexander Ah so this is how you are using LLMs

[Leon @LeonLondon9]

@W98AB SE Ldn?

[Leon @LeonLondon9]

@abacaj Worse than gpt-4 for roughly same cost

[Leon @LeonLondon9]

@StuartReid1929 Gpt-4 best ime, including vs finbert

[Leon @LeonLondon9]

@macrocephalopod Curious to hear more on that, what does integrating quant techniques mean there? You plug in your earnings estimates and model spits out expected return?

[Leon @LeonLondon9]

@6garry9 @ThePythonQuant Now adjust for risk

[Leon @LeonLondon9]

@CasinoCapital 55k is not pretty well as a single person in London...

[Leon @LeonLondon9]

@StuartReid1929 Would recommend you apply to azure openai, there's an application process so takes a while..

[Leon @LeonLondon9]

@realrapxaj @Tobeech Calculate EV and anything below that you re roll

[Leon @LeonLondon9]

@Tobeech Multi managers were hiring like crazy at the start of this year, for sure doable

[Leon @LeonLondon9]

@OkayEstimator Agree that it should never be in the backtesting engine but you often find in the signal itself, most likely created during feature engineering.

[Leon @LeonLondon9]

@0xLoris Ravenpack