Available Mixes

Crosshatch offers several mixes optimized for different use cases and languages. Here's an overview of the currently available mixes:

LMSys Coding Leaderboard Mix

This mix is designed for coding tasks and is based on the top models ranked in the LMSys Chatbot Arena for the Coding category. It automatically adjusts to use the best-performing models, ensuring you always have access to top-tier coding assistance.

Composition

Model NameWeight %

chatgpt-4o-latest

22.32%

claude-3-5-sonnet-20240620

26.61%

gpt-4o-2024-05-13

30.58%

gemini-1.5-pro-exp-0827

20.49%

Update Frequency: Based on LMSys leaderboard changes

LMSys Overall Leaderboard Mix

This mix is based on the top models ranked in the LMSys Chatbot Arena for the Overall category, making it suitable for a wide range of tasks.

Composition

Model NameWeight %

chatgpt-4o-latest

22.51%

gemini-1.5-pro-exp-0827

21.54%

gemini-1.5-pro-exp-0801

23.79%

gpt-4o-2024-05-13

32.15%

Update Frequency: Based on LMSys leaderboard changes

SynthCode Mix

This synthesis mix uses a mixture-of-agents architecture to provide high-quality coding answers. It employs two "proposer" models and an "aggregation" model to synthesize and refine responses.

Composition

Model NameType

Claude 3.5 Sonnet

Proposer

GPT-4 Turbo

Proposer

GPT-4o

Aggregator

Performance

  • 18% better performance than the current leader in Bigcodebench Instruct Hard

  • Scored 31.1% (Pass@1) on 148 problems, compared to 26.4% for GPT-4

Note: This mix may have longer response times due to its complex architecture.

SEAL Spanish Mix

This mix is optimized for Spanish language tasks, based on the top models ranked by Scale.ai's SEAL leaderboard for Spanish.

Composition

Model NameWeight %

gpt-4o

22.51%

gemini-1.5-pro

21.54%

gpt-4-turbo-preview

23.79%

Last updated: est. May 2024

Future Mixes

Crosshatch is continually developing new mixes to address diverse user needs. Some potential future mixes include:

  1. Multilingual Mixes: Optimized for various languages and multilingual tasks.

  2. Role-Playing Mixes: Designed for creative writing and character-based interactions.

  3. Vision Mixes: Combining text and image analysis capabilities.

  4. Speech Mixes: Optimized for speech recognition and generation tasks.

  5. Image Generation Mixes: For creating and manipulating images based on text prompts.

These upcoming mixes aim to expand Crosshatch's capabilities across different domains and modalities. Stay tuned for updates on new mix releases and their specific features and performance metrics.

Last updated