LoRA on Qwen3-VL-2B-Instruct trained via Direct Preference Optimization on 2{,}000 Bradley-Terry preference pairs. No reward model in the loop at training time. Captions emitted directly with no thinking trace.

Training data

271 New Yorker contests, top-rated caption per contest (yguooo/newyorker_caption_ranking).
The 60k Bradley-Terry preference pairs underlying the reward model (separate split).
We deliberately do NOT use the dataset's GPT-4o-generated Scene/Twist/Location/Entities descriptions in the prompt, since they hand-feed scene content to a vision-language model that can already see the image; this makes the policy and reward model usable on any single-panel cartoon, not just the curated subset.

How it fits the project

Part of a 2x2 ablation over training method (SFT, GRPO) and output format (no thinking, thinking) for humor caption generation. See HumorR1/rm-qwen25vl-3b-nodesc for the reward model used to train (and score) this policy.

Inference

Backbone: Qwen/Qwen3-VL-2B-Instruct. This repo is a LoRA adapter; load with peft.PeftModel.from_pretrained.

python

from PIL import Image
from transformers import AutoProcessor
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-2B-Instruct", trust_remote_code=True)
llm = LLM(model="Qwen/Qwen3-VL-2B-Instruct", trust_remote_code=True, dtype="bfloat16",
          enable_lora=True, max_lora_rank=32, max_model_len=4096)

# Caption format: <caption>X</caption>; thinking variant prefixes <think>...</think>.

Reward model used during training

HumorR1/rm-qwen25vl-3b-nodesc (held-out pairwise accuracy 0.6635).

HumorR1/policy-e3-dpo-no-thinking

humor-r1 — DPO, no thinking (Qwen3-VL-2B-Instruct + LoRA) (E3)

Training data

How it fits the project

Inference

Reward model used during training

HumorR1/
policy-e3-dpo-no-thinking