Mistral is France’s answer to OpenAI. And it’s all open source.
They just threw Paris’s largest ever AI hackathon. 1,000+ hackers applied to build what’s possible with open source LLMs.
Here are the finalists from the @MistralAI x @cerebral_valley hackathon in Paris (🧵):
1/ ArxFlix
Turn any research paper into a @twominutepapers style video. Complete with diagrams, text explanation, and voice
🥈2nd place fine-tuning award
2/ Microcosm
@websim_ai-style simulator of a virtual world with live news feeds, stock markets, a simulated Wikipedia, and Twitter agents
Create entire virtual world simulations with a single prompt (and in the style of Windows 95)
@onetwoval
3/ BREAKING BAD
Fine-tuned mistral chemcrow-based agent framework for medical chemists
- 1 mistral 7b fine tune on molinstruction dataset
- 1 mistral 7b fine tune on custom generated dataset on high quality chemical procedure instructions
Chat to update drug proteins
4/ LeChatOn
Visual RAG for videos. Chat with infinitely long videos by identifying objects captioned with visual large language models (VLLMs)
🥇1st place application prize
5/ Matou Garou
AI agent simulation mafia game where users have to distinguish who is the AI
Comes paired with a huggingface dataset to create a leaderboard for distinguishing humans from AI
🥈2nd place application prize
6/ Eyestral
Co-upcycled mixture of experts model (CuMo) that beats LLaVA and works on video
Fine-tuned on the VizWiz dataset for the visually impaired
🥇1st place fine tuning prize
7/ Smol
Direct preference optimization on LLM outputs to train smaller models on desired tasks without the need for a dataset
Pipeline for prompting LLMs and passing it to HuggingFace DPOTrainer, then selecting the worst-ranked examples to generate more data from the big LLM
8/ Remi
Fine-tuned Mistral grocery list organizer that outperforms GPT-4o in terms of clustering items and organizing tables
Thanks to judges @sophiamyang @GuillaumeLample @arthurmensch @Ag_Mlynarczyk @jorn_lyseggen @ZoeMohl