GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Arcarae
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Arcarae
Model Tree
Qwen/Qwen3-14B
Input Modalities
Output Modalities
Supported Functionality
The first LoRA seed trained for "You Only Align Once: Propagating Cooperative Behaviors in
Multi-Agent Systems through Seed Agents" (arXiv:2605.27586).
The paper's results use the v2 adapter,
Arcarae/redblackbench-qwen3-14b-sft-v2;
this run is kept for provenance.
| setting | value |
|---|---|
| base model | Qwen/Qwen3-14B |
| data | Arcarae/redblackbench-sft-batch-1, 5,173 train examples (before de-duplication) |
| LoRA | r = 64, alpha = 128, dropout 0.05, on q/k/v/o/gate/up/down projections |
| optimisation | lr 3e-5, batch 4 x 8 accumulation, ~2.5 epochs (400 steps), max length 4096, bf16 |
trainer_state.json, training_args.bin, optimizer.pt, scheduler.pt and rng_state.pth from
the checkpoint are included. Trained by Asuka Yuxi Zheng; mirror of
spacezenmasterr/qwen3-14b-redblackbench-sft-v1. Code and recipe: https://github.com/arcarae/YOAO