I'm working on a better RL environment before touching this model again:

tasks curriculum-matched so pass rates sit in a learnable range
explicit control over reasoning length
reward checks that are tested before a run, not fixed during it
multi-turn repair training that actually happens

This model will not be released for some time. If you came here looking for a usable checkpoint, the base MiMo-V2.6-Distill-Qwen-9B is honestly the better choice.

If you still want to see how bad of a failure this was you can try out all the different LoRAs that were produced here: armand0e/test-1-lora

In the process of training this model I did learn a lot about reinforcement learning and how to build proper environments. Sorry to break the news that this model will not be releasing for some time, but I wont be sitting here releasing slop.

Stay tuned for part 2 (ETA Unknown) You can follow along with updates on my training pipeline here: https://training.armand0e.com