HumorR1/policy-e1b-sft-thinking

humor-r1 — SFT, with thinking (Qwen3-VL-2B-Thinking + LoRA, merged) (E1b)

LoRA-adapted Qwen3-VL-2B-Thinking supervised fine-tuned on (image, synthetic thinking, chosen caption) triples, then merged. Output format: {thinking}</think>\n\n<caption>X</caption>.

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-Thinking. This repo is a merged full model; load with transformers.AutoModelForCausalLM.from_pretrained.

from PIL import Image
from transformers import AutoProcessor
from vllm import LLM, SamplingParams


processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-2B-Thinking", trust_remote_code=True)
llm = LLM(model="Qwen/Qwen3-VL-2B-Thinking", trust_remote_code=True, dtype="bfloat16",
          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).