Taylor AI

Fine-tune open-source LLMs in minutes

134 followers

Fine-tune open-source LLMs in minutes

134 followers

•

•

Fine-tune open-source LLMs (including Llama2, Falcon, etc) in minutes. With Taylor AI, you can focus on experimentation & building better models, not on digging through python libraries or keeping up with every open-source LLM. And you get to own your models.

Free Options

Launch tags:Developer Tools•Artificial Intelligence•No-Code

Launch Team

LightfieldAI-native CRM that builds itself and does work for you

Promoted

Taylor AI

Maker

📌

Hey Product Hunt! 👋 This is Ben & Brian, and we’re excited to announce Taylor AI! 🧶 Taylor AI (YC S23) empowers enterprises to start fine-tuning open-source LLMs in seconds, so that data science and engineering teams can focus on building great products instead of worrying about GPUs, debugging Python libraries, and keeping up with every new LLM. We started Taylor AI because we saw companies rushing to adopt AI, but struggling to integrate one-size-fits-all chat models like GPT-4 with their proprietary data. Fine-tuning is a great way to customize models, but much of the tooling available for fine-tuning models today is buggy and hard to use. We want to make fine-tuning accessible to every developer and data scientist. With Taylor, you can: 🚀 Start a training run in seconds 👀 Fine-tune state-of-the-art open-source LLMs 🔧 Own your model 🧪 Benefit from cutting-edge techniques (QLoRA, sequence packing, etc.) 🔎 Focus on experimentation, not squashing Python bugs We’re excited for you to try Taylor AI out—every new user can launch up to 3 training jobs for free. Let us know if you have any questions or concerns by emailing us at contact@trytaylor.ai or using the chat on our website (that goes straight to us, not a bot!). What are you fine-tuning LLMs for? Comment below 👇

Report

3yr ago

It's absolutely fantastic! 🔥 Everything bundled together, from fine tuning to quantization and then deployment. I was somehow curious about the quantization process? Could you please specify the method you're using to quantize the model?

Report

3yr ago

Taylor AI

Maker

@zakria Depending on the model, we either use GGML to quantize it to 4-bit, or Ctranslate2 to quantize to 8-bit.

Report

3yr ago

Hey there. Why should I use this Taylor instead of H2O? Cheers.

Report

3yr ago

Taylor AI

Maker

Hey @lyondhur - H2O LLM Studio is a great offering, but it serves a different purpose. H2O requires you to set up your own GPU infra, and you eat those costs if your training run doesn't work. You have to make lots of choices about your training setup, and it's easy to make a small error and mess it up. We handle provisioning compute and training just works out of the box. We're purpose-built and solely focused on fine-tuning. Lots of startups are building in fine-tuning as an afterthought, for us, it's our core product—which is why we had Llama-2 fine-tuning up and running the day after the model was released. We're also pretty focused on teams and collaboration. If you use LLM Studio, all your runs are siloed on your VM. We're building Taylor AI to be collaboration-first, so that you can share your models with your team and work together to train the best one to solve your problem. Finally, we're offering some advanced features not supported by many of our competitors, such as multi-turn conversational AI fine-tuning, which is crucial for enterprise customers that want to build chat models on proprietary data. Happy to chat further on this :)

Report

3yr ago

@brian_kim15 cheers for the reply. Just one correction, you do not need to use GPU at all with H2O if you don't want to, and the setup is pretty trivial. Best of luck with the project.

Report

3yr ago

Taylor AI

Maker

@brian_kim15 @lyondhur We've talked to a lot of users who want something different than what H2O offers. We're excited to build a solution that allows more people to get started training models to solve their problems fast.

Report

3yr ago

Hey @benjamin_anderson5 @brian_kim15 ! While testing, I encountered a minor bug, so I uploaded an incorrect dataset file, causing the program to get stuck in a loop. There's currently no means to halt or terminate the process. It would be beneficial if you guys could incorporate a stop button or a similar feature to interrupt the process and during the training process too.

Report

3yr ago

Taylor AI

Maker

@brian_kim15 @zakria Thanks for pointing this out! We're working on more informative error messages & failing more gracefully when there's a problem. There have also been some issues in the last 24h due to a Cloudflare outage with our hosting provider. We're learning from these bumps in the road and the experience will definitely get smoother over time.

Report

3yr ago