Show HN: We built the Keras of LLMs | Hacker News Reader

Show HN: We built the Keras of LLMs

github.com

5 pointsyoansallami2 comments

Here is a production-first Keras-inspired LM framework, built with the advice of François Chollet (ex-Google, creator of Keras and ARC-AGI), our technical advisor. This system have already been deployed in production with our clients (which is why we have already every LLMOps practice implemented). It is also compatible with Jupyter and Marimo to integrate seamlessly in you Data Scientists workflows. You can try the code examples online on HF space and you can find more information in the documentation and FAQ. If you have any feedback for us don't hesitate to join our discord! More releases are planned in the next months and over the year. Have a great day!

Open articleView on HN

Keras is built with a Functional API at its core. How it works is that it infer the whole pipeline specification (the tensor shapes) and construct a DAG (directed acyclic graph) of trainable layers only by knowing the inputs and outputs, in our case the tensor are replaced with a JsonDataModel with a JSON schema for specification. In this version, each layer (called Module) is a trainable element that can be optimized using In-Context Reinforcement Learning (and other RL algorithms in the near future). Because there is no gradient descent but only a RL-like back-propagation we can have inactive branches and build logical flows, with python operators like `|` (or) and `&` (and), on top on them. Which make easier to build any workflow or Agent with it. Synalinks also optimize with asyncio each module call thanks to the DAG. We use constrained structured output by default, and combined with in-context RL we can ensure a correct syntax, and content at the same time. In addition to that, the application is serializable in JSON and the best performing program (with associated prompts) is automatically saved and can be versioned with git. You can try it online, you will find more explanation and diagrams: https://huggingface.co/spaces/YoanSallami/synalinks-notebooo...

Very interesting! Could you elaborate a bit on the parallels with Keras, other than the API similarities? What aspects of working with LLMs does SynaLinks make easier?

Reply on news.ycombinator.com