Open-sourced jev architecture last year with model,paper and dataset

97 points•nandakishor_ml•14 days ago•18 comments•

#x2f #paper #dataset #jev #architecture #https #sourced #x27

Everyone now talks about the architecture that's not auto regressive and does lightning fast probability prediction with a json schema. I worked on this literally one year back in March 2025, published an arxiv paper, pushed the model to huggingface along with the pypi package and training dataset. And then one year later, a frontier lab came, proposing the same idea like literal breakthrough without technical papers, open weights and no open dataset. For anyones information the main guiding model is RL not embedding model or LLM

Paper: https://arxiv.org/abs/2503.23303

Model: https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning

Dataset: https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations

Also the second work published in September 2025 was exactly the same one jev proposed now

Paper: https://arxiv.org/abs/2510.01237

My model uses PPO over sequence embeddings to output turn-by-turn conversion trajectories (probabilities from 0.0 to 1.0).

Jev uses parallel sampling (trained via RLCD) to output confidence distributions and schema choices.

It's incredibly frustrating that the thing that you made with months of hard work, sweat and sleepless night is architecturally similar with the vertical use case and don't get the support you deserve because frontier lab build something horizontal. The open-source story in general

18 comments

aesthesia12 days ago

From what I can tell these are classifiers trained for a single task. What excites people about Jev is that it can do zero-shot structured responses for arbitrary prompts. Now, this isn't new either; models like GLiNER have been around for a while. But Jev appears significantly more flexible and polished while still being cheap and fast.

gokuljs11 days ago

Such interesting comments. I’d never find this kind of discussion on Twitter. It makes me think

tomrod12 days ago

> It's incredibly frustrating that the thing that you made with months of hard work, sweat and sleepless night is architecturally similar with the vertical use case and don't get the support you deserve because frontier lab build something horizontal.

Totally get the frustration.

Don't get too bent up about it -- take it as the market validating your hypothesis in a way you didn't expect. You should feel proud! That skillset - finding niches that could be humongous given the right cultivation, luck, funding, and marketing - is amazing.

I feel strongly that, if JEV and/or your model prove valuable as many of us are already thinking and hypothesizing, people will come knocking.

While LLMs are a neat space, this is a desperately missing component, and as someone who has built conditional choice probability models for almost two decades, I'm wildly interested to dig in this weekend and review the usefulness, the applicability as a homo economicus level of automation in a noisy prompt space (e.g. ensuring transitivity and IIA), and looking at this as a strong evolution.

I've added your model to my eval list!

creamyhorror14 days ago

The world's heavily about marketing, resources, connections, and signaling, unfortunately. You probably needed to market it in a bigger forum with shinier claims to attract attention (I don't think a paper on Arxiv is enough).

BSVogler8 days ago

I tested your model quickly and it seems that the model does not do what is claimed here (that it can do the same as jev). It relies on an LLM, while jev does not need an LLM. While it is interesting and novel work it is not a replacement for expensive LLM calls.

For anyone interested in playing around with it, it requires python 3.11, else it will segfault.

According to the code the conversation turn index is a signal that enters so it is judging on something that does not transfer to other tasks. In the sampling it moved the score more than the actual content.

8MEM6 days ago

Really interesting setup. Given the recent interest in non-autoregressive fast probability prediction, do you plan on extending your framework into a more horizontal, general-purpose open-source model?

Read the full thread on Hacker News →

18 comments

Related stories