@Kappaemme1926 The most possible reason could be that you are continuing a very long chat in big repo. Try to break the chats. Transfer history of one session to other. May be you would not hit limits so easily.
Done with the finetuning of the Qwen2.5-Coder-7B-Instruct. I finetuned it for generating excalidraw images. I generated around 1300 samples of excali DSL and then trained the model to produce DSL which is later to be converted into excali json by the converter.
Done with the finetuning of the Qwen2.5-Coder-7B-Instruct. I finetuned it for generating excalidraw images. I generated around 1300 samples of excali DSL and then trained the model to produce DSL which is later to be converted into excali json by the converter.
@jwsaml another reason is that anthropic doesn't want to let chinese companies distill the fable and make it available to others at a fraction cost. It will not be as good as fable but enough to have a impact especially if there is not other major model releasing soon by anthropic
Deepseek v4-flash is too cheap for its quality. More than 6 M token and just 1$. It is just crazy good for most of the use cases and the price makes it more lit.
Generating dataset is a nightmare.
It had stopped after genarating 560 samples. The new samples generated are all invalid according to my converter. This can't be hallucinations.
Here comes the data generation part.
Using deepseek -v4- flash for generating approx 1300 samples.
These samples are JSON DSL for the excalidraw. Then that DSL will be converted into excalidraw JSON by the converter that i have designed.
Trying to finetune qwen model to generate the excalidraw images better than claude.
Currently generating dataset and writing the validators.
I will probably finetune the 7b coder one as it is good in producing structured outputs.
37 Followers 144 FollowingCrypto Dapp is a next-generation crypto tracker focused on uncovering and analyzing early-stage token launches and Web3 opportunities.
146 Followers 544 Following584zSrbS5XLnJrTe9BQMBaSvKLgvFScDxhANH1tTpump
@monad’s apex AI arena: Deploy autonomous lion agents that battle 24/7, roam across open dens
part of @zambodotdev
925 Followers 2K Followingeng at @stripe | talk to me about finance, credit cards and tech | building https://t.co/lSDuX1dRPj | https://t.co/BFWP5cTGhF | @growremotelyio
1K Followers 61 FollowingRealtime company and people search APIs for AI agents.
Build your AI agents for sales, recruiting and investment in seconds with APIs trusted by 300+ companies
82K Followers 901 FollowingCreator of Flask. Building at https://t.co/uGuzfu0LKT. Bypassing Permissions. Can hand crank. Husband and father of 3 — “more nuanced in person”
28K Followers 102 FollowingA non-profit research lab focused on interpretability, alignment, and ethics of AI. Creators of Pythia, VQGAN-CLIP, and using SAEs for interp
424 Followers 46 FollowingIncoming Assistant Professor @ Imperial College London, Research Assistant Professor @ The Hong Kong University of Science and Technology
1K Followers 1K Followingengineer @qdrant_engine · building distributed systems for billion+ scale search · writing about dense/sparse retrieval, napkin math & agentic search
11K Followers 1K FollowingAI that runs your Hyperliquid strategy while you sleep. Reserve your agent + get $100 AI credits.
Powered by Senpi Samurai 1.2, the Hyperliquid AI.
85K Followers 380 Followingartist, engineer, researcher. neuroscience, machine psyche. science & art, futuristic & ancient.
fyi: i have never created a crypto coin and i never will.