RakshithFury/Maatu-15m-base-2b
Text Generation • 15M • Updated
I would not have tried for a thinking mode/COT at such low token scales and yes, IFT,RL can only amplify the ability in the base model so if the base model isnt capable enough, doing RL will not help much.
I have tried RL only on larger models, but if the base sub 500m model is very capable i do not see why RL should not boost.
Are you specifically trying to build models at low budget/token scales?
I see. Makes sense. Next intuitive step would be RL.
So the model has 2 stages - Pretraining for Chat data SFT or Instruction tuning
@jiteshkprasad yes, i cannot expect a model trained on only 18B tokens to be good at math, code or any reasoning intensive task
Have you done any RL like GRPO on it?
Also are you using RULER- Niah for the niah numbers?