OpenPipe linked up w/ Wyatt Marshall CTO & Co-Founder of Halluminate so he could have an in-depth conversation on how to build a robust Evals system for your production GenAI technology w/ Reid Mayo (Founding AI Engineer).
Check it out!:
OpenPipe
116 posts
OpenPipe
@OpenPipeAI
OpenPipe: Fine-tuning for production apps. Train higher quality, faster models. (YC S23)
- Our CEO @corbtt is giving a talk next Wed in SF w/ @neutralino1 and @charles_irl. Kyle's topic is on fine-tuning best practices to get far higher quality LLM results vs prompting alone, and when to explore this optimization technique. check it out! lu.ma/jfsyz0xc
- We <3 Helicone5 tips to slash your LLM costs: 1️⃣ Optimize your prompts 2️⃣ Implement response caching (use @helicone_ai to help you save $$ and time) 3️⃣ Use task-specific, smaller models (We recommend @OpenPipeAI for this, you can fine-tune a cheaper model that can result in as much as an 85%
- We've added support for fine-tuning GPT-3.5-0125, OpenAI's newest and cheapest model. Sign up and check it out! 🙂 @OpenAIDevs
- PSA because apparently we're getting confused with closed-down competitors: all OpenPipe users can export their fine-tuned models at any time. No lock-in nonsense here!Replying to @5harath and @eter_inquirer@UnslothAI for if you know Python and have a GPU or can set up your own training on @modal_labs (or otherwise). Unsloth is super fast and easy. If you need a full platform - @OpenPipeAI could be a good fit. I dont think you get the raw model back though now since they moved to
- Curious about Llama 2? Here's a fun feature we shipped last week: automatically convert your GPT-3.5 prompt to the Llama 2 format with best practices! Play with Llama 2, Claude 2, and GPT models at openpipe.ai 😉
GIF - Great walkthrough of how to use OpenPipe to create state-of-the-art fine-tuned models!
- Super pumped. 💪Automatically link Helicone with OpenPipe to finetune models to improve inference speed, cost, and accuracy. Sign up for our beta here: helicone.ai/integration/op…
- Replying to @tcarambat @5harath and 3 othersWe 100% do let you export your raw model and do whatever you want with it, and will never remove that capability. Super aligned with the open source ecosystem. 🙂
- I am not sure whether we want more Sydneys in the world but if you do you can 100% use OpenPipe to build them. 🫡
- Available on huggingface ofc, and also on the OpenPipe platform!Just dropped a new Mistral variant "Mistral 7B Fine-Tune Optimized." Built to be a very strong base for downstream fine-tuning. On our evals fine-tunes trained on it are stronger than fine-tunes on any other 7B model, and even beat gpt-4. 🙂 openpipe.ai/blog/mistral-7…
- I've now heard from two different sources that character.ai is fine-tuning hyperpersonalized LoRAs on a per-character and possibly per-user basis. I expect we'll see much more of this! LoRAs aren't much more expensive to serve than base models.
- Replying to @DamiDina
