Deploy selfplay trainer from b3202c405e19f6b4f9defecc748ed0358111eef2

Deploy selfplay trainer from 9062ef625693685dfe748a2444f2c6baea9eb8ea

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Deploy selfplay trainer from 312af1c2877300e47381e4ca80a195f5906fe623

Deploy selfplay trainer from e7d94bf79fc5c66fbd4b3b11607712a4c86ef753

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Deploy selfplay trainer from d3356ff330ec4809bb60f3fb8e35595b8e2ea282

Local sync: darkguard-selfplay-trainer

Local sync: darkguard-selfplay-trainer

Deploy selfplay trainer from 18ea3b212e9bcaa7f61732b6d91eca80594678c4

Sync darkguard-selfplay-trainer from RL_Env (curriculum + fixes)

Rebalance designer challenge_delta toward challenging but solvable

Avoid consumer penalty double-counting in reward routing

Remove deprecated Blocks theme arg for Gradio 6

Remove deprecated wandb start_method setting

Wire router toggle into policy runtime

Disable heavy local router by default and clean generation args

Expose local action router checkbox in GUI

Add local action router toggle default off

Set default base models to unsloth Qwen3 FP8

Add local Qwen-based action routing layer with safe fallback

Wire env retry/pacing config into trainer client

Add connection retry and pacing config defaults

Add rate-limit pacing and retry backoff for env requests

Preserve secret tokens when GUI fields blank and auto-enable W&B

Improve W&B auth fallback and init reliability

Read HF/W&B tokens from Space secrets by default

Log rollout warnings instead of fatal crash

Handle reset/step failures without crashing trainer

Improve remote env HTTP error diagnostics

Fix connection/start callback errors with safe parsing

Always clear running state via trainer finally block

Recover stale running state on Start/Stop

Improve stop button UX when idle

Allow holdout evaluation to stop early

Fix stop button responsiveness in phase loops

Update GUI default model to unsloth Qwen3 4B FP8

Set default base models to unsloth Qwen3 4B FP8

Initial self-play trainer Space sync