Qwen3.5-0.8B-Picture-Captioner

A lightweight vision-language model fine-tuned to generate concise, detailed, and accurate image captions.

Model Details

• Base Model: Huihui-Qwen3.5-0.8B-abliterated (uncensored)
• Training Dataset: agentlans/monet-caption
• Precision: bf16
• Attention Mechanism: sdpa (Scaled Dot-Product Attention)
• Max Pixels: 768 × 768 = 589,824

Click here for training details

Training Configuration

Parameter	Value
Epochs	1
LoRA Rank	16 (α = 32)
Dropout	0.05
Learning Rate	2e-4
Weight Decay	0.01
Warmup Steps	100

Training Results

• Training Loss (Smoothed): 2.1571
• Validation Loss: 2.2069

Input Format & Usage

• To format your inputs correctly for inference, structure your prompt as follows.
• To save RAM, also resize your images to 768 × 768 proportionally

content = [
    {"type": "image", "image": image},
    {"type": "text", "text": "Describe this image in detail."}
]

# Recommended generation settings
# tokenize = False
# add_generation_prompt = False

Click here to see example images and captions.

Limitations

While efficient and capable, this model shares common failure modes typical of smaller Vision-Language Models (VLMs):

• Object and Action Misidentification: May occasionally hallucinate or misidentify specific objects (e.g., mislabelling a red ball as a tomato) or unusual actions (e.g., a person pushing a penguin into a shopping cart).
• Repetitive Captions: Despite training data variety, the model can occasionally fall into repetitive phrasing loops.
• Inference Speed: As a VLM, generation speeds are inherently slower than traditional text-only models or specialized tagging methods.

Licence

This model is licensed under the Apache 2.0 Licence.