Hi everyone, and thanks for following ChenkinNoob-XL. This post brings you up to speed on where we are with releases and what we’re planning—including why 0.3 was delayed and why we’re moving to 0.5.

1. Why wasn’t 0.3 released on time?

We didn’t ship 0.3 as planned for one main reason: The results didn’t meet our own bar, so we didn’t feel right releasing it.

We tried higher dropout and more aggressive repeat strategies to strengthen character quality, but the model kept degrading. After fixing dropout and repeat, quality was still getting worse, so we spent a long time debugging. We only recently confirmed the main cause: the e621 dataset. Once we removed it, quality improved clearly and we’re back on track.

That experience reinforced our approach: we’d rather take the time to get a release right than ship an unsatisfactory 0.3 just to hit a date. Thank you for your patience and for keeping us on our toes—we’ll keep pushing from here and aim to deliver a 0.5 that lives up to your expectations.

2. Why are we jumping from 0.3 to 0.5?

We’ve decided to call the next release 0.5 instead of 0.3, for two reasons.

(1) Training is much faster—we can move quicker

We’ve overhauled our training pipeline; under the same resources, training speed is now about 3× faster. That lets us iterate faster without cutting quality, and move more quickly toward 1.0. We’re reflecting that in the version number (0.3 → 0.5): this isn’t a small patch, but a **clear step up in pace**, with the goal of reaching 1.0 sooner and more steadily.

(2) Refreshing data and version together

The data used for the 0.3-era runs was cut off at 2026-01-07; it’s now March, so we have over two months of new data to fold in. We plan to retrain on this updated set to improve recency and diversity. With that larger data refresh, we’re bumping the version to 0.5 to match the scope of the update.

(3)What about the “0.3” number?

We’re keeping 0.3 as the label for that exploratory phase (fixing e621, tuning dropout/repeat). 0.5 is the new starting point—the first full-sized release after those fixes. That way we honor the work and expectations around 0.3, while making the 0.5 name meaningful.

---

In short: **0.5 = back on track after fixing e621 (and related issues), plus “faster training” and “updated data”**—a solid step toward 1.0.

Before 0.5: Chenkin RF v0.3

Before we ship 0.5, the CKN ecosystem will release Chenkin RF v0.3. It uses the same **Danbooru data cutoff** as the 0.3 run—**no e621**—with more training steps than 0.2, and under the Rectified Flow setup it performs better than our half-finished 0.3. Think of it as a bridge release until 0.5—we’d love for you to try it and share feedback.