theblackcat102/Molmo2-4B-Pairwise-Judge-Direct-Augmented

Molmo2-4B Pairwise Judge (Direct Augmented, merged)

This is a standalone BF16 Molmo2-4B model fine-tuned for pairwise judging of AI-generated videos. Given a user's preference questionnaire, a generation prompt, and two candidate videos, it emits exactly one token: 1, 2, or Tie.

The model was produced by merging the LoRA adapter from molmo2_pairwise_judge_direct_augmented into the local Molmo2-4B base. The adapter used rank 32, alpha 64, dropout 0.05, and targeted att_proj, attn_out, ff_proj, and ff_out.

Loading

The repository contains the Molmo2 custom model, processor, tokenizer, and video-processing code needed for trust_remote_code=True:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "theblackcat102/Molmo2-4B-Pairwise-Judge-Direct-Augmented"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
)

Use the Molmo2 chat-template format with type="video" content. The answer should be constrained or post-processed to the expected labels (1, 2, or Tie) when evaluating this judge.

Prompt format

Send one user message containing the preference questionnaire, the generation task and prompt, then the two candidate videos in order. Finish with the question asking for exactly one label. The order of the videos defines the meaning of 1 and 2.

messages = [{
    "role": "user",
    "content": [
        {
            "type": "text",
            "text": """You are simulating a specific human judge of AI-generated videos. You are given that person's video-preference questionnaire and two candidate videos generated for the same prompt. Decide which of the two videos this person would prefer, or whether they would consider them a tie.

User video preference questionnaire.
Which video qualities matter most: {qualities}
Most wanted kind of generated video: {wanted_kind}
Preferred amount of motion: {motion_preference}
Preferred camera treatment: {camera_preference}
Preferred overall visual treatment: {visual_preference}
{example_comparisons}""",
        },
        {
            "type": "text",
            "text": "Generation task: {task}.\nPrompt used to generate both videos: \"{generation_prompt}\"",
        },
        {"type": "text", "text": "Video 1:"},
        {"type": "video", "video": "/path/to/video_1.mp4"},
        {"type": "text", "text": "Video 2:"},
        {"type": "video", "video": "/path/to/video_2.mp4"},
        {
            "type": "text",
            "text": "Which video would this person prefer: 1, 2, or Tie? Answer with exactly one token.",
        },
    ],
}]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
)
outputs = model.generate(**inputs, max_new_tokens=3, do_sample=False)
answer = processor.tokenizer.decode(
    outputs[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
).strip()

{example_comparisons} should contain zero or more lines in the same style as the training data, for example: Example comparison - City flight (option A: Model A; option B: Model B): Preferred option A, preference strength: Moderate, decision factors: Camera and framing. Keep the final request unchanged and pass videos in the same order as their labels.

Provenance and validation

The base model was the local allenai/Molmo2-4B export at merge time. The merged weights contain no PEFT adapter wrapper and load as Molmo2ForConditionalGeneration. A fresh-process smoke test loaded the standalone model and processor and ran a local two-video validation example, generating 2 with greedy decoding and three new tokens.

This model is intended for research evaluation of video preferences. Its outputs reflect the training judge and should not be interpreted as an objective measure of video quality.