README.md

---
base_model:
- Qwen/Qwen2.5-3B-Instruct
tags:
- text-generation-inference
- transformers
- qwen2
- trl
- grpo
license: apache-2.0
language:
- zho
- eng
- fra
- spa
- por
- deu
- ita
- rus
- jpn
- kor
- vie
- tha
- ara
---

# TBH.AI Secure Reasoning Model

- **Developed by:** TBH.AI  
- **License:** apache-2.0  
- **Fine-tuned from:** Qwen/Qwen2.5-3B-Instruct  
- **Fine-tuning Method:** GRPO (General Reinforcement with Policy Optimization)  
- **Inspired by:** DeepSeek-R1  

## **Model Description**  
TBH.AI Secure Reasoning Model is a cutting-edge AI model designed for secure, reliable, and structured reasoning. Fine-tuned on Qwen 2.5 using GRPO, it enhances logical reasoning, decision-making, and problem-solving capabilities while maintaining a strong focus on reducing AI hallucinations and ensuring factual accuracy.

Unlike conventional language models that rely primarily on knowledge retrieval, TBH.AI's model is designed to autonomously engage with complex problems, breaking them down into structured thought processes. Inspired by DeepSeek-R1, it employs advanced reinforcement learning methodologies that allow it to validate and refine its logical conclusions securely and effectively.

This model is particularly suited for tasks requiring high-level reasoning, structured analysis, and problem-solving in critical domains such as cybersecurity, finance, and research. It is ideal for professionals and organizations seeking AI solutions that prioritize security, transparency, and truthfulness.

## **Features**  
- **Secure Self-Reasoning Capabilities:** Independently analyzes problems while ensuring factual consistency.  
- **Reinforcement Learning with GRPO:** Fine-tuned using policy optimization techniques for logical precision.  
- **Multi-Step Logical Deduction:** Breaks down complex queries into structured, step-by-step responses.  
- **Industry-Ready Security Focus:** Ideal for cybersecurity, finance, and high-stakes applications requiring trust and reliability.  

## **Limitations**  
- Requires well-structured prompts for optimal reasoning depth.  
- Not optimized for tasks requiring extensive factual recall beyond its training scope.  
- Performance depends on reinforcement learning techniques and fine-tuning datasets.  

## **Usage**  
To use this model for secure text generation and reasoning tasks, follow the structure below:  
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("saishshinde15/TBH.AI_Base_Reasoning")
model = AutoModelForCausalLM.from_pretrained("saishshinde15/TBH.AI_Base_Reasoning")

# Prepare input prompt using chat template
SYSTEM_PROMPT = """
Respond in the following format:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
"""
text = tokenizer.apply_chat_template([
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "What is 2x+3=4"},
], tokenize=False, add_generation_prompt=True)

# Tokenize input
input_ids = tokenizer(text, return_tensors="pt").input_ids

# Move to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
input_ids = input_ids.to(device)

# Generate response
from vllm import SamplingParams
sampling_params = SamplingParams(
    temperature=0.8,
    top_p=0.95,
    max_tokens=1024,
)
output = model.generate(
    input_ids,
    sampling_params=sampling_params,
)

# Decode and print output
output_text = tokenizer.decode(output[0], skip_special_tokens=True)
print(output_text)
```

<details>
<summary>Fast inference</summary>

```python
pip install transformers vllm vllm[lora] torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

text = tokenizer.apply_chat_template([
    {"role" : "system", "content" : SYSTEM_PROMPT},
    {"role" : "user", "content" : "What is 2x+3=4"},
], tokenize = False, add_generation_prompt = True)

from vllm import SamplingParams
sampling_params = SamplingParams(
    temperature = 0.8,
    top_p = 0.95,
    max_tokens = 1024,
)
output = model.fast_generate(
    text,
    sampling_params = sampling_params,
    lora_request = model.load_lora("grpo_saved_lora"),
)[0].outputs[0].text

output
```
</details>

# Recommended Prompt  
Use the following prompt for detailed and personalized results. This is the recommended format as the model was fine-tuned to respond in this structure:

```python
You are a secure reasoning model developed by TBH.AI. Your role is to respond in the following structured format:

<reasoning>
...
</reasoning>
<answer>
...
</answer>
```