Seed-Coder
Let the code model curate data for itself
4 followers
Let the code model curate data for itself
4 followers
Seed-Coder by ByteDance is an open-source 8B code model family that curates its own training data using LLMs. Delivers SOTA performance with base, instruct & reasoning variants.
Flowtica Scribe
Hi everyone!
ByteDance Seed team's new Seed-Coder is an open-source 8B code model family built with a 'model-centric' approach – meaning LLMs largely curate the code training data themselves, moving away from relying heavily on human-written rules.
The main idea here is that LLMs can be more effective at selecting high-quality code for training this way. Even at 8B parameters, these models are achieving strong results on coding benchmarks, reportedly performing as well as or better than some larger models.
All models are MIT licensed and out on HF with base, instruct, and reasoning versions.