README.md

---
license: apache-2.0
tags:
- RakutenAI
- DeepSeek-R1
- task-vector-merging
- japanese
- multilingual
library_name: transformers
language:
- ja
- en
pipeline_tag: text-generation
base_model:
- deepseek-ai/DeepSeek-R1-0528
- Rakuten/RakutenAI-3.0
- deepseek-ai/DeepSeek-V3-0324
---
# RAI-3.0-R1-VECTOR
<a href="https://www.apache.org/licenses/LICENSE-2.0" target="_blank" style="margin: 2px;">
  <img alt="License" src="https://img.shields.io/badge/License-Apache%202.0-4caf50?&color=4caf50" style="display: inline-block; vertical-align: middle;"/>
</a>

---

## Model Overview
**RAI-3.0-R1-VECTOR** is a task-vector merged model created using the following formula:
```
DeepSeek-R1-0528 + (RakutenAI-3.0 - DeepSeek-V3-0324)
```

This architecture combines the advanced reasoning capabilities of `DeepSeek-R1-0528` with the Japanese language expertise of `RakutenAI-3.0`, while subtracting the base `DeepSeek-V3-0324` to isolate task-specific improvements.

## Key Features
- **Enhanced Reasoning**: Inherits DeepSeek-R1's improved depth of reasoning (average 23K tokens per complex query).
- **Japanese Optimization**: Retains RakutenAI-3.0's proficiency in Japanese language and cultural context.
- **Reduced Hallucination**: Benefits from DeepSeek-R1's reduced hallucination rate.
- **Multilingual Support**: Balanced performance in both Japanese and English.

## Technical Details
| Parameter                | Value                          |
|--------------------------|--------------------------------|
| Base Model               | DeepSeek-R1-0528              |
| Task Vector Source       | RakutenAI-3.0 - DeepSeek-V3-0324 |
| Architecture             | Mixture of Experts (MoE)      |
| Context Length           | 128K tokens                   |
| License                  | Apache-2.0                    |

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Local-Novel-LLM-project/RAI-3.0-R1-VECTOR", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Local-Novel-LLM-project/RAI-3.0-R1-VECTOR")

inputs = tokenizer("日本の文化で重要な要素は", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))
```

## Limitations and Bias
- May inherit biases from either source model.
- Performance in non-Japanese/English languages may vary.
- Always verify critical outputs with human review.

## Citation
```bibtex
@misc{RAIR1VECTOR2026,
  title = {RAI-3.0-R1-VECTOR: Task-Vector Merged Model},
  author = {LocalNovelLLM-project},
  year = {2026},
  publisher = {LocalNovelLLM-project},
  url = {https://huggingface.co/Local-Novel-LLM-project/RAI-3.0-R1-VECTOR}
}
```

Note: This model card was generated by the model itself and subsequently edited.