hari

retired supercard player✌lets see how long it lasts

[hari @hari94809472]

Despite all criticisms this was still a free pcc and personally one of my best

[hari @hari94809472]

@altudev @AhmetAyyldz_ @Zai_org How does it compare to gpt models? Anything better than gpt 5.4xhigh-5.5 models and i might just take it.The limits are atrocious at the 20$ range for the frontier model subscriptions

[hari @hari94809472]

@Teknium Is this that's something already there/can it be implemented?

[hari @hari94809472]

@Teknium Basically doing grep search to find relevant tools through file structure

[hari @hari94809472]

@jun_song I have read papers on this but the general consensus with constrained decoding is when you restrict the logits the reasoning of llm takes a hit also i don't think we have a neat agent harness that supports constrained decoding.I see outlines and stuff but that just api calls

[hari @hari94809472]

@jun_song my understanding is it's tough to modify as it's mostly pretraining issue or you postrain but I'm not sure what /how you postrain and fix the reward model.Any resources/links for me to better understand the approach you are using?Does constrained decoding play a part to fix this?

[hari @hari94809472]

@jun_song Hi Jun ,
Day 3 of asking you how as a developer one might go about fixing tool calling errors in models?im unable to find much info online and would love to know how you solve those issues, keenly waiting for your reply

[hari @hari94809472]

@harshbhatt7585 Hi harsh,
Been curious to know if you might have any idea how tool calling is generally fixed by the providers? Anything a developer can do to fix them on release with/without gpu access?can't find much info online on this and given your background im open to ideas

[hari @hari94809472]

@jun_song Hey Jun sorry for asking again I'm an AI developer and noticed your massive contribution in the field,I have before asked you how you have been able to solve tool calling issues in models,as an AI developer it would really be helpful as less information is available online on it

[hari @hari94809472]

@jun_song @k3ntosan Hey Jun i wanted to ask you how exactly people are fixing tool calls exactly?I don't see my ch relevant info online and it's a problem i constantly face with local models

[hari @hari94809472]

@jun_song @waefrebeorn How exactly does one fix tool calling if you don't mind me asking im more interested in the technical details and how generally constrained decoding is fixed in practice?

[hari @hari94809472]

@bnjmn_marie @marcaruel Btw were there any tests you did that checked tool calling capabilities of gemma?

[hari @hari94809472]

@bnjmn_marie @marcaruel Im not a bit.i used to play this game I just follow ai a lot now I'm a ai developer as a professional

[hari @hari94809472]

@bnjmn_marie Which versions are you testing and what is the actual model id you are using

[hari @hari94809472]

@bnjmn_marie Fixed by Google or independent publishers?

[hari @hari94809472]

@laman_gh @Bhnewman05 @Hesamation Im More interested in the structured output side of things.I don't see many implementations online where you can pair an agent with structured output generation other than through json.Internally they use context free grammar to implement them for json they still fail occasions

[hari @hari94809472]

@laman_gh @Bhnewman05 @Hesamation Link?

[hari @hari94809472]

@SarikaMaha86540 @ShivaKarka @mindmusclepro Lab tests, brand matters,periodic testing with appropriate measures to not bias the test etc, channels like trustified are doing this