Looking for collaborators (dataset curation / fine-tuning) — ZYR3.1MoE, open-weights multi-agent orchestration model

Beginners
1

Hi, I’m the developer behind ZYR3.1 (zyr-AGENT/ZYR3.1MoE). It’s an open weights, merged, standalone model — 9B parameters, fully self-contained (no adapter/PEFT step needed), built on a Qwen base.

The design is a layered multi-agent orchestration stack: an ATP layer handles planning, task decomposition, and verification; an ACN layer carries messages between components; and MeO (Router/Reasoner/Synthesizer) decides which of 11 specialist agents — Planner, Researcher, Reasoner, Coder, Math, Critic, Fact Checker, Creative, Optimizer, Reviewer, and a Computer agent — get involved for a given task. Simple questions skip the orchestration entirely and go straight to the base model; only real work requests trigger the full plan → route → build → verify loop.

I’ve also been careful to separate trained behavior from real execution — there’s a demo trace showing the simulated “team chat” style the model produces, and a separate, genuinely externally-verified trace (real commands, real exit codes, real pass/fail) so people can see the difference rather than take marketing copy at face value.

I’m running this solo right now and looking for one or two collaborators who enjoy dataset curation, agent orchestration, or fine-tuning work — especially if you’ve worked on tool-use datasets, agent trajectories, or MoE merges. Happy to go deep on the architecture or the eval setup with anyone interested.

Model card: zyr-AGENT/ZYR3.1MoE · Hugging Face

2

Model card’s pretty thorough on the trained-vs-real-execution distinction, that’s a good habit most agent-orchestration repos skip.

One thing worth flagging before people jump in: the benchmark table is 20 questions total (5 each across 4 categories), that’s way too small a sample to call “100% overall,” even a couple of easy/leaked questions would explain that score. If you want collaborators to take the orchestration claims seriously, a larger public eval set (even 100-200 questions) would help a lot more than the current table.

For the dataset curation side, worth being specific in your ask, are you after people to help build tool-use/agent-trajectory data for the ATP/MeO routing decisions specifically, or more general instruction-tuning data for the merged 9B base? That’ll matter a lot for who bites, since trajectory/routing data collection is a pretty different skill set from general SFT dataset work.

3

Would YOU like to join my team? for free

??

4

Appreciate the offer, but I’ve got my own solo projects running right now (search engine, tokenizer/ML research work) so I don’t have the bandwidth to properly commit to a second project. Good luck with ZYR3.1 though, the trained-vs-real-execution distinction in the model card is a good sign for finding contributors who take it seriously.