Mithula Artigala PRO

bro, this guy is einstein 😭

now if you still want help, share the training recipe with us

the 8BA1B will be better in real world and that will have interesting things to explore for you.

once again. can you share your training recipe (so I can evaluate what you did wrong)

interesting... can you give me your training recipe?

ohhh absolutely! @Bc-AI

dawg, why did you overcomplicate this. u just had to tell this VERY SIMPLE word: on-policy 😭
anyways, good luck on it, man! I believe you can do it :)

ofc this is assuming ur using the KL divergence to boil down the representations. I mean... what else are you supposed to do for this very specific scenario.

well... lets be serious here.
if you are talking about "distillation" you could either directly or indirectly distill it.

Direct

almost all your models are from different family with their own tokenizers, so you are NOT doing logit level distillation for sure.

indirect distillation

well... this is kind what you are doing. it mainly consists of smart tricks, which consists of: YOU GUESSED IT! SFT or RL😭

you can directly move the models likeliness for the next token to a specific style of tokens with SFT, which is something we all do. do for fun actually.

then you can do RL. for example:

• DPO: you reject the student's sample and use the teacher's sample
• GRPO: you can give the teacher's higher scores. although if the teacher is too big or if there is too much samples from a bigger teacher, it directly attacks point sfrom this research: 2502.12143 (yalls hould read it, its a really cool paper)

also there is this new technique called MOPD, its made for tasks like these. tbh, for distillation, i reccomend you to this paper: arXiv:2502.12143

but anyways, you should take down that statement, bcs you are NOT doing direct distillation with missmatching tokenizers, and you are doing SFT or RL!

its not bad. wouldbe nice if you could have a bit more samples (at least 1k browser use and 4k samples for creative workloads)

also rserve some for whatever RL you do.

I mean... you can use an LLM to search for it. maye there is some data outside HF as well.

can you include browser-use and design work as well?

I would say:

• use open design (or sum like that) for the design harness
• for the browser use harness, use whatever you are cmfortabelwith. just at least more than 1 though.

ohh yeah. I didnt even try it yet but their older 24B model was already good. so i am not expecting nothing less from it!

also LFM makes really solid models.

yes! I am very certain it will be good. dont just train it on agentic coding though. can you use something like the open-design harness and a browser use harness for the data aswell? thanks!

for the 27B model. would be great if you could work o the LFM2.5-24B-A2B model!

First. Kudos to you. This is research and work we absolute need! ❤️

second. by quantizing the KV-cache that aggressively you are going to slow down an already easy-to-run run 1.2B and 2.6B model.

I am certain that many devices can comfortably run a 1.2B model or a ~3B model. if not you could quantize the model and KV cache (to about 8bits) which will allow the model to fit, but besides that, its the speed which will make it obsolete.
How are you going to adress this, @Banaxi-Tech ?

edit: this would be helpful for a 6B+ model though.

yeah, we gotta appreciate it 😭

positional embeddings. sorry about that.