watercolour-grpo-judge-led

A LoRA adapter for Qwen/Qwen3.5-35B-A3B, trained with GRPO to paint watercolours by writing p5.brush sketches. This is the run with the original reward mix from Surya Narreddi's write-up: the pairwise judge, the term that carries the hand-rated reference pool, holds most of the weight. Of the three runs in this project, this one optimises the curator's taste the hardest.

Loading it, because the obvious way fails silently

Qwen3.5-35B-A3B declares Qwen3_5MoeForConditionalGeneration and carries a vision tower, so its layers live at model.language_model.layers. AutoModelForCausalLM resolves to the text-only variant, whose layers sit at model.layers, and 700 of the adapter's 920 tensors then fail to match. PEFT reports that as a UserWarning, not an error, so you get the base model back and nothing tells you.

from transformers import Qwen3_5MoeForConditionalGeneration, AutoTokenizer
from peft import PeftModel

base = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3.5-35B-A3B", dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(base, "HuggingEnvs/watercolour-grpo-judge-led")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-35B-A3B")

If you see Found missing adapter keys while loading the checkpoint, the adapter did not load and you are running the base model.

Training

110 steps in 34h49m on one H200. Launched for 200 steps and stopped at 110, with the curve still inching upward (+0.0013/step over the last 30). This adapter is the step-110 checkpoint.

reward	gate 0.05 + length 0.05 + pairwise judge 0.60 + HPSv3 0.30
learning rate	5e-5, constant_with_warmup, 5 warmup steps
scale_rewards	none
LoRA	all-linear, r16, alpha 32. 30,431,360 trainable, 0.0866%
batch	8 generations per step, per_device_batch_size 1, grad_accum 8
sampling	top_p 0.95, top_k 20, max_completion_length 8192

WATERCOLOUR_JUDGE_WEIGHT=0.60 WATERCOLOUR_QUALITY_WEIGHT=0.30 \
hf jobs uv run examples/watercolour_grpo.py --flavor h200 --timeout 96h --secrets HF_TOKEN -- \
  --env-url https://YOURORG-watercolour-env.hf.space \
  --model Qwen/Qwen3.5-35B-A3B --lora --all-linear --bf16 --gradient-checkpointing \
  --subject 'a peach hibiscus' --references 4 \
  --top-p 0.95 --top-k 20 \
  --lr 5e-5 --lr-scheduler constant_with_warmup --warmup-steps 5 \
  --scale-rewards none \
  --steps 200 --num-generations 8 \
  --per-device-batch-size 1 --gradient-accumulation-steps 8 \
  --max-completion-length 8192 --probe-samples 40 --film

The uv header pins no versions (trl, peft, transformers, torch), so a run today will resolve different ones. That is a real reproducibility gap, stated rather than hidden.

Results

	first third	second	third	slope t
reward	0.451	0.647	0.721	+10.5
pairwise judge term	0.36		0.70
HPSv3 term	0.56		0.77
paint coverage	0.11		0.23

Absolute rewards are not comparable across reward mixes: each run optimises a different blend. What is comparable is the shape: this run started the lowest and climbed the most, spending its first thirty steps nearly flat before it moved.

Best group mean 0.888, at step 53, and the best single rollout of the run is 0.96. The pairwise judge term itself climbed from 0.36 to 0.70: the model wins more comparisons against the hand-rated pool as training advances. Paint coverage doubled, from 0.11 to 0.23, where the judge-free hps-only run barely moved it.

The base model's probe before training: reward 0.407, judge term 0.317, paint coverage 0.093, over 40 samples. Every training number here is recomputable from watercolour-rollouts-judge-led.

Siblings

run	judge	HPSv3
judge-led	0.60	0.30
hps-led	0.30	0.60
hps-only	0.00	0.90

Limitations

• One subject, a peach hibiscus, and one library. It does not generalise to other drawing tasks.
• The pairwise judge is the noisiest reward term, and its consistency (scoring the same image twice) has not been tested.
• Reproducing it needs an H200, an a100-large Space for HPSv3, a cpu-upgrade Space for the environment and inference quota for the judge. It is not cheap.

Method reproduced from Surya Narreddi's "RL'ing Qwen to paint with code". Internally this run is v22b, HF job 6a95460d0718b0f6d8908805.

Where this comes from

Part of Paint with Code, a complete recipe: the environment, the pool that defines the reward, the trainer, the curves and every rollout.

the recipe, and how to reproduce it	02-watercolour/
the environment	envs/watercolour/
the trainer	train/watercolour_grpo.py
the reference pool	watercolour-reference-pool
the trained adapter	watercolour-grpo-judge-led
every rollout	watercolour-rollouts-judge-led