Open Source Today (2024-08-09): Tongyi Qianwen Releases Qwen2-Math for Advanced Math Reasoning

Explore the latest AI open-source projects like Qwen2-Math for advanced math, Parler-TTS for fast TTS, and more innovative tools and frameworks.

Here are some interesting AI open-source models and frameworks I wanted to share today:

Project: Qwen2-Math

Qwen2-Math is a series of large language models in the Qwen2 family, designed specifically for math problems. It comes in three versions: 1.5B, 7B, and 72B parameters.

This model performs exceptionally well in solving arithmetic and mathematical problems, surpassing both open-source and closed-source models like GPT-4.

Qwen2-Math is intended to solve advanced math problems that require complex, multi-step logical reasoning.

https://huggingface.co/Qwen/Qwen2-Math-1.5B

https://huggingface.co/Qwen/Qwen2-Math-7B

https://huggingface.co/Qwen/Qwen2-Math-72B

Project: Parler-TTS

Parler-TTS has released Mini (880M) and Large (2.3B) model weights, trained on 45,000 hours of audiobook data. Compared to version 0.1, the generation speed is 4 times faster.

It also supports SDPA and Flash Attention 2 for even more speed.

Parler-TTS is a lightweight text-to-speech (TTS) model that generates high-quality, natural-sounding speech in the style of specific speakers (gender, tone, speaking style, etc.).

https://github.com/huggingface/parler-tts

Project: MoonPalace

MoonPalace is an API debugging tool from Moonshot AI. It works on Mac, Windows, and Linux. It’s easy to use—just replace the base_url with http://localhost:9988 after launching to start debugging.

This tool captures full requests, including detailed information during network errors. You can quickly retrieve and view request details using request_id and chatcmpl_id.

MoonPalace also allows one-click export of structured BadCase data to help improve the Kimi large model's capabilities.

https://github.com/MoonshotAI/moonpalace

Project: LiteMultiAgent

LiteMultiAgent is a library for LLM Agent applications.

The project aims to improve the efficiency of multi-agent systems by classifying and parallelizing agents.

It organizes agents into a hierarchy by categorizing different tool sets, allowing for more types of tasks. Sub-agents act as tools and execute tasks in parallel naturally.

https://github.com/PathOnAI/LiteMultiAgent

Project: RPBench-Auto

RPBench-Auto is an automated pipeline for evaluating the performance of large language models in role-playing tasks.

The project supports multiple model configurations and generates performance rankings through a series of predefined evaluation tasks.

Users can run evaluations and submit results to a public leaderboard with simple command-line operations.

https://github.com/boson-ai/RPBench-Auto

Project: AutoGGUF

AutoGGUF provides a graphical user interface for quantizing GGUF models using the llama.cpp library.

It allows users to download different versions of llama.cpp, manage multiple backends and perform quantization tasks with various options.

https://github.com/leafspark/AutoGGUF