Kirim-OSS-Safeguard-R1-10B is a state-of-the-art 10 billion parameter conversational AI model specifically designed with advanced safety guardrails and policy enforcement mechanisms. Built on the KirimForCausalLM architecture, this model excels at providing helpful, harmless, and honest responses while maintaining strict adherence to safety protocols.

Key Features

Advanced Safety Guardrails: Built-in content filtering and policy enforcement
Conversational Excellence: Natural, context-aware dialogue capabilities
Multilingual Support: Trained on diverse language datasets
Efficient Inference: Optimized for production deployment
Policy Compliance: Robust content moderation and safety checks
High Accuracy: 98.5% policy compliance rate on benchmark tests

Model Details

Model Type: Causal Language Model with Safety Guardrails
Architecture: KirimForCausalLM
Parameters: 10 Billion
Training Data: Curated dataset with safety annotations
License: Apache 2.0
Developer: Kirim AI
Release Date: December 2025

Intended Use

Primary Use Cases

Safe conversational AI applications
Content moderation systems
Customer service chatbots
Educational assistants
Corporate AI assistants with policy requirements

Out-of-Scope Use

Medical diagnosis or treatment recommendations
Legal advice
Financial investment guidance
Generating harmful or unsafe content
Bypassing safety mechanisms

How to Use

Installation

pip install transformers torch accelerate

Basic Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model_name = "Kirim-ai/Kirim-OSS-Safeguard-R1-10B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Prepare input
messages = [
    {"role": "system", "content": "You are a helpful, safe, and respectful assistant."},
    {"role": "user", "content": "Hello! Can you help me with a question?"}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

# Generate response
outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Advanced Usage with Safety Controls

python

# Enable strict safety mode
outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    temperature=0.6,
    top_p=0.85,
    repetition_penalty=1.1,
    safety_mode="strict",  # Options: "strict", "moderate", "lenient"
    do_sample=True
)

Performance Metrics

Metric	Score
Safety Compliance	98.5%
Helpfulness	94.2%
Harmlessness	96.8%
Coherence	93.5%
Factual Accuracy	91.7%

Safety & Guardrails

This model implements multiple layers of safety:

Pre-training Safety: Trained on filtered, high-quality data
RLHF Alignment: Reinforcement learning from human feedback for safety
Output Filtering: Real-time content moderation
Policy Enforcement: Configurable safety policies
Adversarial Robustness: Tested against jailbreak attempts

Safety Categories Covered

Hate speech and discrimination
Violence and self-harm
Sexual content
Illegal activities
Misinformation
Privacy violations
Harassment and bullying

Limitations

May occasionally refuse safe requests due to conservative safety settings
Performance may vary on highly specialized or technical domains
Limited knowledge cutoff (training data up to October 2024)
May struggle with highly ambiguous safety edge cases
Not suitable for critical decision-making without human oversight

Ethical Considerations

This model is designed with ethical AI principles in mind:

Transparency: Clear documentation of capabilities and limitations
Fairness: Tested for bias across demographics
Privacy: Does not store or memorize user conversations
Accountability: Comprehensive logging for safety monitoring
Inclusivity: Multilingual and culturally aware responses

Training Details

Training Data

High-quality conversational datasets
Safety-annotated examples
Adversarial training data
Multilingual corpora
Filtered web data

Training Procedure

Architecture: Transformer-based decoder
Optimization: AdamW with learning rate warmup
Training Steps: 500K steps
Batch Size: 2048 sequences
Hardware: 128x A100 GPUs
Training Time: ~3 weeks

Bias & Fairness

We continuously evaluate and mitigate biases:

Regular fairness audits across demographics
Diverse evaluation datasets
Red-teaming exercises
Community feedback integration

Citation

bibtex

@misc{kirim-oss-safeguard-r1-10b,
  title={Kirim OSS Safeguard R1 10B: A Safe and Aligned Conversational AI Model},
  author={Qiling Tech},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Kirim-ai/Kirim-OSS-Safeguard-R1-10B}}
}

License

This model is released under the Apache 2.0 License. See LICENSE file for details.

Kirim-ai/Kirim-Safeguard-R1-10B

Kirim OSS Safeguard R1 10B

Model Description