[ARCHIVE]2026-07-16T18:00:49.396371+00:00
Qwen3 4b Instruct from 1% to 99% on wordle in 11 hours with a single 5090 with PPO.
Executive Summary
Hello everyone! I’ve previously worked on transformers and ppo for traditional rl environments but I wanted to see if my training would scale up to llm fine tuning (it does at least for wordle). I modified/rewrote my gridworld agents from mapox-trainer to load Qwen3 weights and created a new value a...
Deep analysis unavailable for this source.
View Original SourceClassification: Open