OpenPipe

RLHF-curious? I’ve put together a very practical guide to building a task-specific reward model! Includes lots tips on choosing the right metric and data, and all code is included. Hope it’s helpful. 🙂 If your application has human feedback (regenerations, user choices, etc.) please DM me and I’d love to chat about how we can use RLHF to improve your response quality significantly with minimal marginal effort!

Yesterday two new major model families became available for fine-tuning: Llama 3.1, which comes in 8B, 70B and 405B(!) variants, and GPT-4o mini. We’ve added them to the OpenPipe platform and ran all of them (except Llama 3.1 405B) through our evaluation harness. The good news is, all 3 of models are extremely high quality. The bad news is, they saturate most of the standard evals we ran, which makes comparing them difficult! In fact, both Llama 3.1 variants we tried saturate all 3 of the standard evals we ran, and GPT-4o mini also saturated 2/3 of them. What do we mean by saturate? For any given input, you can imagine there is a potential “perfect” output (or set of outputs) that cannot be improved upon. The more complex the task, the more difficult it is for a model to generate a perfect output. However, once a model is strong enough to consistently generate a perfect output for that task, we consider the task saturated for that model. In our LLM-as-judge evals, this usually shows up as a cluster of models all doing about the same on the task without any model significantly outperforming. And in fact, that's what we see in the evaluations below. All 3 fine-tuned models do about as well as each other (win rates within 6%) on both the "Resume Summarization" and "Data Extraction" tasks. On "Chatbot Responses" however, both Llama 3.1 variants significantly outperform GPT-4o mini. So the "Chatbot Responses" task isn’t saturated for GPT-4o mini, but all other tasks and models are. This is very significant—we chose these tasks explicitly because older models on our platform, like Mistral 7B and Llama 3 8B, did not saturate these tasks! There are two main reasons why we’re seeing this saturation now: - The new models we’re testing here are stronger than the previous generation of models available on-platform. - Our benchmark models are now all trained on datasets relabeled with Mixture of Agents, which substantially improves the quality of the dataset and thus the fine-tuned model. We’re working on developing better benchmarks, and once we have some higher-difficulty ones we’ll analyze Llama 3.1 405B as well. And again, you can try all these out today on OpenPipe to run your own evaluations!

🔥 The AI Tinkerers - Seattle Summer Hackathon is ON! 🔥 👉 https://lnkd.in/g7QJgTxZ This event brings together two of the most interesting and innovative early stage AI companies -- Moondream & OpenPipe -- together with the AI Tinkerers community to imagine what's possible to open source generative AI models through fine-tuning and vision. We invite you to bring your team, or form a team at the event. Judges, prizes, food and founders of those two companies will be on hand. June 15th. 🔥 Huge thanks to the generous sponsorship of Ascend.vc 🔥 Kirby Winfield Jen Haller Kyle Corbitt David Corbitt Jason Allen, Vik

Here is a sample math explainer video that me and my team at the #Llama3Hackathon generated this past weekend: https://lnkd.in/gBN4U63n I fine-tuned #Llama3 8B on generating the code utilizing the Manim python library to create videos based on user inputted math questions using OpenPipe and performed inference on OctoAI. My dataset came from code generated from Gemini 1.5. You can find our project here: https://math.auto.movie  Blog description of the project: https://lnkd.in/gqmJiFNM  Github repo: https://lnkd.in/gnXH7hwW

OpenPipe (YC S23) has raised $6.7M in seed funding to replace GPT-4 with your own fine-tuned models. Founded by Kyle Corbitt and David Corbitt, OpenPipe makes it easy to replace your GPT-4 prompt with a fine-tuned model. This is especially beneficial for projects or businesses making over 1000 OpenAI prompt calls a day, as it offers improvements in speed, cost, and quality without compromises. If you have a prompt already in use, OpenPipe can help you collect and refine your data, train a new model, and deploy it quickly, often in less than an hour, even for those without machine learning experience. Once your fine-tuned model is operational, OpenPipe enhances its value by automating the process of improving your model with growing user data, a concept known as a data flywheel. This strategy, used by major companies like Amazon for superior product recommendations, allows your model to improve continuously as your user base expands. OpenPipe enables any company to begin implementing machine learning workflows and building a sustainable competitive edge by leveraging early data for ongoing enhancement. Congrats on the round, Kyle and David!