Ruurd/byod-llama-3.1-8b

BYOD-Llama-3.1-8B

Interactive full-precision inference for BYOD-Llama-3.1-8B, one of the BYOD (Build Your Own DLM) models. It uses the exact best LoRA checkpoint from the corresponding experiment and loads its original base model in BF16. No 4-bit quantization is used.

Generation starts from masked answer positions and refines them in parallel. Fewer denoising steps than generated tokens provide a sub-autoregressive inference budget; increasing the step count gives the model more refinement opportunities.

Model: Ruurd/BYOD-Llama-3.1-8B

This is a research demo. Outputs may be inaccurate or inappropriate and inherit limitations from the original base model.