sergiopaniego/rl-envs-youtube-livestream-4-scripts

Updated 27 days ago

Note The training scripts, the eval script and the launcher, with the configs that produced the numbers

Note The OpenEnv environment: MBPP problems inside a live Python session

Note One run, mbpp-grpo: flat train/reward (t=1.41) next to a rising eval/reward (t=5.18), same weights

sergiopaniego/qwen3-1.7b-mbpp-grpo

Text Generation • 2B • Updated 27 days ago • 477

Note Demo 1, white box: GRPOTrainer + environment_factory. 0.518 to 0.582 on 257 unseen problems

sergiopaniego/Qwen3-8B-opencode-deepcoder-grpo

Text Generation • 8B • Updated Jul 28 • 47 • 1

Note Demo 2, black box: the opencode agent driving its own loop in remote sandboxes

Note The task data for demo 1

Note The task data for demo 2

Note Dashboard for demo 2: the reward climbing from 0.27 to 0.71 over 10 steps

Note The extra not shown in class: the same white-box hookup against a text game

sergiopaniego/qwen3-1.7b-wordle-grpo

Text Generation • 2B • Updated Aug 19 • 59

Note The wordle extra, trained with train_wordle_whitebox.py