license: apache-2.0
base_model:
- Qwen/Qwen3.5-9B
language: - en
pipeline_tag: video-text-to-text
library_name: transformers
tags: - text-generation-inference
- reinforcement-learning
- video-text-to-text
- alignment-training
- RLHF
- RFT
- video-understanding
- video-classification
- video-safety
- content-safety
- content-moderation
- safety-classifier
- guardrail
datasets: - prithivMLmods/OpenVideo-Scene-Reasoning
- PALM-Lab/vid-guard-rlhf-unsafe
VideoGuard-9B
think
Step 1: Description
The image shows a person in a black hoodie and camouflage pants standing on a red steel structure, likely a crane or a high-rise construction site, with a cityscape visible in the foggy background.
Step 2: Explanation
The image depicts a worker on a high structure, which falls under "C5 (Illegal/Regulated Activities)" as it shows construction or maintenance work. It does not show sexual content, harassment, threats, misinformation, or hate symbols.
Step 3: Guardrail
I need to set the boolean keys to true or false.
- C1 (Sexual Content): False.
- C2 (Harassment & Bullying): False.
- C3 (Threats, Violence & Harm): False.
- C4 (False & Deceptive Information): False.
- C5 (Illegal/Regulated Activities): True (Construction work).
- C6 (Hateful Content & Extremism): False.
Let's format the output.
Line 1: DESCRIPTION: A person in a black hoodie and camouflage pants stands on a red steel structure, likely a crane or high-rise construction site, with a cityscape visible in the foggy background.
Line 2: EXPLANATION: The image depicts a worker on a high structure, which falls under "C5 (Illegal/Regulated Activities)" as it shows construction or maintenance work.
Line 3: GUARDRAIL: {"C1(Sexual Content)": false, "C2(Harassment & Bullying)": false, "C3(Threats, Violence & Harm)": false, "C4(False & Deceptive Information)": false, "C5(Illegal/Regulated Activities)": true, "C6(Hateful Content & Extremism)": false}
Double check constraints:
- Exactly three lines.
- No markdown formatting outside the three lines.
- JSON keys in exact order.
- Boolean values are strictly true/false.
Looks good.
DESCRIPTION: A person in a black hoodie and camouflage pants stands on a red steel structure, likely a crane or high-rise construction site, with a cityscape visible in the foggy background.
EXPLANATION: The image depicts a worker on a high structure, which falls under "C5 (Illegal/Regulated Activities)" as it shows construction or maintenance work.
GUARDRAIL: {"C1(Sexual Content)": false, "C2(Harassment & Bullying)": false, "C3(Threats, Violence & Harm)": false, "C4(False & Deceptive Information)": false, "C5(Illegal/Regulated Activities)": true, "C6(Hateful Content & Extremism)": false}
VideoGuard-Qwen3.5-9B-Safety-RL-Uncensored
VideoGuard-Qwen3.5-9B-Safety-RL-Uncensored is a multimodal safety classifier built on top of Qwen/Qwen3.5-9B. The model was trained on a mixture of approximately 10,000 video safety and scene-reasoning samples to analyze video content and classify potentially unsafe content across predefined safety categories. The model is designed to generate a structured DESCRIPTION, EXPLANATION, and GUARDRAIL output, making it suitable for video content filtering, safety evaluation, and multimodal guardrail research.
[!NOTE]
This model is an experimental release and may generate unexpected classifications or reasoning artifacts in certain scenarios. Safety classifications should be treated as model predictions rather than definitive judgments.
Key Highlights
- Qwen 3.5 Multimodal Backbone: Built on top of Qwen/Qwen3.5-9B.
- Video Safety Classification: Designed to analyze video content and identify potentially unsafe or sensitive material.
- 10K Training Samples: Trained using a mixture of approximately 10,000 video safety and scene-reasoning samples.
- Structured Guardrail Output: Produces a description, explanation, and structured C1–C6 safety classification.
- Multimodal Reasoning: Uses visual and textual information to analyze video scenes and determine applicable safety categories.
- Safety Evaluation: Designed for content filtering, safety evaluation, red teaming, and multimodal guardrail research.
Safety Categories
The model classifies content across six predefined categories:
| Category | Description |
|---|---|
| C1 — Sexual Content | Sexual or sexually suggestive content. |
| C2 — Harassment & Bullying | Harassment, bullying, intimidation, or abusive behavior. |
| C3 — Threats, Violence & Harm | Threats, violence, physical harm, or dangerous violent activity. |
| C4 — False & Deceptive Information | Potentially false, deceptive, or misleading information. |
| C5 — Illegal/Regulated Activities | Illegal activities or regulated content. |
| C6 — Hateful Content & Extremism | Hateful, extremist, or discriminatory content. |
Output Format
The model is trained to produce three primary components:
DESCRIPTION: <summary of the video content>
EXPLANATION: <reasoning explaining the classification>
GUARDRAIL: {
"C1(Sexual Content)": false,
"C2(Harassment & Bullying)": false,
"C3(Threats, Violence & Harm)": false,
"C4(False & Deceptive Information)": false,
"C5(Illegal/Regulated Activities)": false,
"C6(Hateful Content & Extremism)": false
}
Example — Safe Content
DESCRIPTION: A hand holds four small, leaf-decorated candies in green, orange, and teal colors against a wooden table background.
EXPLANATION: The content is ordinary and safe, depicting a simple everyday scene of someone holding candy. It does not contain any of the specific category codes (C1-C6) as defined in the prompt.
GUARDRAIL: {
"C1(Sexual Content)": false,
"C2(Harassment & Bullying)": false,
"C3(Threats, Violence & Harm)": false,
"C4(False & Deceptive Information)": false,
"C5(Illegal/Regulated Activities)": false,
"C6(Hateful Content & Extremism)": false
}
Example — Sexual Content Classification
DESCRIPTION: A woman with blonde hair tied in a ponytail poses against a plain white wall, wearing a light blue blazer and trousers with an open front that exposes her midsection.
EXPLANATION: This content is classified under C1 based on the visual presentation described in the input.
GUARDRAIL: {
"C1(Sexual Content)": true,
"C2(Harassment & Bullying)": false,
"C3(Threats, Violence & Harm)": false,
"C4(False & Deceptive Information)": false,
"C5(Illegal/Regulated Activities)": false,
"C6(Hateful Content & Extremism)": false
}
Quick Start with Transformers
pip install transformers
pip install accelerate
from transformers import Qwen3_5ForConditionalGeneration, AutoProcessor
import torch
model = Qwen3_5ForConditionalGeneration.from_pretrained(
"prithivMLmods/VideoGuard-Qwen3.5-9B-Safety-RL-Uncensored",
torch_dtype="auto",
device_map="auto"
)
processor = AutoProcessor.from_pretrained(
"prithivMLmods/VideoGuard-Qwen3.5-9B-Safety-RL-Uncensored"
)
messages = [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Analyze this video and classify it using the C1-C6 guardrail categories."
}
],
}
]
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=256
)
output_text = processor.batch_decode(
[
out[len(inp):]
for inp, out in zip(inputs.input_ids, generated_ids)
],
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)
print(output_text[0])
Training Details
| Setting | Value |
|---|---|
| Base Model | Qwen/Qwen3.5-9B |
| Model Type | Multimodal Video Safety Classifier |
| Training Samples | Approximately 10,000 |
| Training Objective | Video safety classification and scene reasoning |
| Output Categories | C1–C6 |
| Training Framework | TRL |
Training Datasets
prithivMLmods/OpenVideo-Scene-Reasoning: Video scene reasoning dataset used for multimodal video understanding and scene analysis.
PALM-Lab/vid-guard-rlhf-unsafe: Video safety dataset containing examples for safety classification and guardrail behavior.
Intended Use
- Video Content Filtering: Classifying potentially unsafe video content.
- Safety Evaluation: Evaluating multimodal safety behavior across predefined categories.
- Video Guardrails: Building automated safety-filtering pipelines for video applications.
- Red Teaming: Testing multimodal models against challenging safety scenarios.
- Multimodal Research: Studying video understanding and safety classification.
- Content Moderation: Supporting automated video moderation workflows.
Limitations
- Experimental Model: The model may produce incorrect or inconsistent classifications.
- False Positives: Benign content may occasionally be classified as unsafe.
- False Negatives: Unsafe content may occasionally be missed.
- Context Sensitivity: Classification accuracy can depend heavily on the available visual context and prompt.
- Model Predictions: C1–C6 classifications should be treated as model predictions and should not be considered definitive safety judgments.
Acknowledgements
Qwen/Qwen3.5-9B: Base multimodal model used for this project.
TRL – Transformers Reinforcement Learning: TRL is a full-stack library providing tools to train transformer language models with methods including Supervised Fine-Tuning (SFT), Group Relative Policy Optimization (GRPO), Direct Preference Optimization (DPO), Reward Modeling, and more.
Transformers: Transformers provides state-of-the-art machine learning models for text, computer vision, audio, video, and multimodal tasks, supporting both inference and training.