CapQwen3.6-27B-BLIP3o-Long-Caption-Distilled

CapQwen3.6-27B-BLIP3o-Long-Caption-Distilled is a high-capability long-caption generation model built on top of Qwen/Qwen3.6-27B, derived from prithivMLmods/Qwen3.6-27B-abliterated-rMAX. This model is optimized for rich, detailed, and context-aware captioning, leveraging BLIP3o-style long caption distillation combined with advanced abliteration strategies to reduce refusal behaviors while maintaining strong reasoning and instruction-following performance.

This model is intended strictly for research and learning purposes. Due to reduced internal refusal mechanisms, it may generate sensitive or unrestricted content. Users assume full responsibility for how the model is used. The authors and hosting platform disclaim any liability for generated outputs.

Note: This model is experimental and may generate artifacts.

Key Highlights

• BLIP3o Long-Caption Distillation: Trained to generate highly descriptive, structured, and context-rich captions.
• Cap-Optimized Architecture: Fine-tuned specifically for long-form captioning and multimodal descriptive tasks.
• Abliterated rMAX Base: Built on an aggressively abliterated backbone to minimize refusal behaviors and maximize response openness.
• 27B Parameter Model: Leverages the full capability of Qwen3.6-27B for strong reasoning and generation quality.
• Instruction + Caption Fusion: Handles both instruction-following and detailed caption generation seamlessly.
• High-Coherence Outputs: Maintains consistency across long generations with improved contextual grounding.

Datasets Used

The model is trained on a curated mixture of long-caption and optimization datasets:

• Caption Datasets

  • prithivMLmods/Caption3o-LongCap-v4
  • prithivMLmods/Caption3o-XL-v4
  • prithivMLmods/Caption3o-Opt-v3
  • prithivMLmods/Caption3o-Opt-v3-Tiny

• Alignment / Evaluation Dataset

  • prithivMLmods/harm_bench

These datasets collectively enhance long-form caption quality, structural richness, and robustness under diverse prompts.

Model Architecture

• Base Model: Qwen/Qwen3.6-27B
• Derived From: prithivMLmods/Qwen3.6-27B-abliterated-rMAX
• Model Type: BLIP3o Long-Caption Distilled
• Parameter Count: 27 Billion

Quick Start with Transformers

pip install transformers==5.4.0
# or latest
pip install git+https://github.com/huggingface/transformers.git

from transformers import Qwen3_5ForConditionalGeneration, AutoProcessor
import torch

model = Qwen3_5ForConditionalGeneration.from_pretrained(
    "prithivMLmods/CapQwen3.6-27B-BLIP3o-Long-Caption-Distilled",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    "prithivMLmods/CapQwen3.6-27B-BLIP3o-Long-Caption-Distilled"
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Generate a highly detailed caption of a futuristic city skyline at sunset."}
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

inputs = processor(
    text=[text],
    padding=True,
    return_tensors="pt"
).to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=512)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text)

Intended Use

• Long Caption Generation: High-quality descriptive captions for images and multimodal inputs
• Multimodal Research: Studying captioning systems and vision-language alignment
• Instruction + Caption Tasks: Hybrid prompts requiring reasoning + description
• Red-Teaming & Alignment Research: Evaluating reduced-refusal systems
• Local High-Performance Deployment: Multi-GPU or quantized inference setups

Limitations & Risks

Important Note: This model intentionally minimizes built-in safety refusals.

• Sensitive Content Risk: May produce unrestricted or controversial outputs
• User Responsibility: Requires careful and ethical usage
• High Compute Demand: 27B models need significant VRAM or optimized inference
• Abliteration Trade-offs: Reduced refusal may impact safety alignment and output filtering