factor out nn_train.py into a separate python project to make it more manageable, add plotting and offline strength fit

1 files changed, 12 insertions, 0 deletions

diff --git a/train/README.md b/train/README.md
new file mode 100644
index 0000000..53f803f
--- /dev/null
+++ b/train/README.md
@@ -0,0 +1,12 @@
+# ctak-train
+
+Self-play training loop for the ctak evaluators.
+
+```sh
+cd train
+uv run nn-train --size 5 --iters 100 --games 200
+```
+
+Each iteration writes `../data/selfplay-<size>-v<iter>.json`, writes candidate
+and champion weights to `../include/`, refits the candidate strengths over the
+whole game history and redraws `../data/candidate_strength.png`.