Base model: Qwen/Qwen3-VL-4B-Instruct
Adapter: LoRA, rank 32, alpha 32 (task_type: CAUSAL_LM)
Algorithm: GRPO (verl), group size 16, seed 42, ~500 steps
Prompt / reward: 80% recursive-self-aggregation prompts, one-shot. Shaped OVEN reward (format 0.05, exact/fuzzy 0.70, specificity-weighted hF 0.15, path-match 0.05, aggregation 0.05); a missing \boxed{} answer returns 0.00.

Training dynamics (from the wandb run)

metric	start	end	notes
policy entropy	0.630	0.071	min 0.058
KL to reference	0.001	7.840	max 8.809
training reward	0.216	0.238	—
response length	162.961	164.918	tokens
val exact-match@1	0.140	0.138	held-out

Training reward moved 0.216 → 0.238 while held-out exact-match went 0.140 → 0.138 (essentially flat). Consistent with the thesis: outcome-only GRPO does not raise validation accuracy above the prompt-only elicitation floor. As a small-KL LoRA update the policy stays close to the reference model.

Usage

python

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True, device_map="auto")
model = PeftModel.from_pretrained(base, "jucamohedano/qwen3-vl-4b-oven-grpo-aggregation")
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct", trust_remote_code=True)

Intended use

Research artifact for reproducing the thesis's (largely negative) reinforcement-learning results on OVEN and for studying GRPO training dynamics on a multimodal open-world task. Not tuned or recommended for production classification.

Provenance

MSc thesis, University of Trento (Juan Camacho Mohedano). Training: verl (branch grpo-oven-v080); reward in verl/utils/reward_score/oven_boxed.py; data by oven-mllm-eval/scripts/build_verl_oven_parquet.py. wandb run: offline-run-20260707_094828-qwen3-vl-4b-oven-grpo-agg08-one-shot-3k-shaped-n16-seed42.