HuggingEnvs/watercolour-grpo-hps-only

Reinforcement Learning • Updated about 6 hours ago • 16

Note GRPO adapter for Qwen3.5-35B-A3B. Reward is 0.90 HPSv3 with the pairwise judge off.

Note The 178 paintings that define the reward, with the sketch that produced each one.

Note Every rollout of the HPS-only run: 470 paintings, sketches and rewards, by step.

HuggingEnvs/watercolour-grpo-judge-led

Reinforcement Learning • Updated about 7 hours ago

Note GRPO adapter for Qwen3.5-35B-A3B. The original mix: pairwise judge at 0.60, HPSv3 at 0.30.

HuggingEnvs/watercolour-grpo-hps-led

Reinforcement Learning • Updated about 7 hours ago

Note GRPO adapter for Qwen3.5-35B-A3B. The middle mix: HPSv3 at 0.60, pairwise judge at 0.30.

Note Every rollout of the judge-led run: 861 paintings, sketches and rewards, by step.

Note Every rollout of the hps-led run: 872 paintings, sketches and rewards, by step.

Note The OpenEnv environment: renders each sketch and computes the reward. Duplicate it to train.

Note The HPSv3 scorer behind the quality term. Duplicate it on a100-large for a run.

Note Live training curves of the judge-led run, per step and per metric.

Note Live training curves of the hps-led run, per step and per metric.

Note Live training curves of the hps-only run, per step and per metric.