Index-1.9B is a series of lightweight, multilingual Large Language Models (LLMs) developed by Bilibili, designed for efficient performance across various tasks including chat, role-playing, and long-context processing. Targeting researchers and developers, it offers competitive benchmarks against similarly sized models and notable multilingual capabilities, particularly in East Asian languages.

How It Works

The Index-1.9B series is built on a 1.9 billion parameter architecture, pre-trained on a 2.8T corpus with a significant portion of Chinese and English data. Variants include a base model, a pure control version, a chat-aligned model using SFT and DPO, a RAG-enhanced character model for role-playing, and a 32K context length version. This approach allows for a compact model with strong performance and specialized capabilities like extended context handling.

Quick Start & Requirements

• Install: pip install -r requirements.txt
• Prerequisites: Python, Transformers library. GPU/MPS acceleration is supported but not strictly required.
• Demo: Web demo (python demo/web_demo.py), CLI demo (python demo/cli_demo.py), and OpenAI-compatible API (python demo/openai_demo.py) are available.
• Long Context: The Index-1.9B-32K model requires a specific tool: demo/cli_long_text_demo.py.
• Quantization: Requires bitsandbytes==0.43.0 for 4-bit quantization.
• Links: HuggingFace, ModelScope

Highlighted Details

• Achieves competitive benchmark scores against models like Phi-2 and Qwen1.5-1.8B.
• Index-1.9B-32K variant supports a 32K context window, outperforming similarly sized models in long-context tasks.
• Offers specialized models for chat, role-playing (with RAG), and base functionality.
• Supports 4-bit quantization for reduced memory usage.

Maintenance & Community

The project is actively maintained by Bilibili. Community interaction is facilitated via a QQ group.

Licensing & Compatibility

• Source Code: Apache-2.0 license.
• Model Weights: INDEX_MODEL_LICENSE. Fully open for academic research and free commercial use.

Limitations & Caveats

The model may generate inaccurate or biased content and cannot express personal opinions. Users are advised to use generated content cautiously and conduct safety tests before deployment. The developers disclaim responsibility for misuse or issues arising from model usage.