Fine-tuning recipe
QLoRA training on the pi_agent split of Fable-5 traces targets agentic behavior and tool use.
MODEL & EVALUATION / GITHUB
A model experiment you can inspect from data to evaluation.
Fabletron brings Nemotron fine-tuning, exported weights, and evaluation tooling into one inspectable project.
THE PROJECT, IN PRACTICE
What you can do with it.
Published weight formats
Choose among LoRA, BF16, GGUF, NVFP4, and MLX artifacts for different research and runtime paths.
Compare with the base
The evaluation harness reports differences against the base across instruction-following, reasoning, coding, and function-calling benchmarks.
FROM IDEA TO USE
How it works.
- 01
Choose a format
Match the model artifact to the runtime and hardware you intend to use.
- 02
Inspect the model card
Review architecture requirements, provenance, and licensing before download.
- 03
Run and evaluate
Use the documented serving path and compare results for your own workloads.
LoRA Apply the fine-tune delta
BF16 Full-precision merged weights
GGUF Compatible llama.cpp / LM Studio
NVFP4 NVIDIA Blackwell serving
MLX Apple Silicon workflowsRuntime support and memory requirements differ. See each artifact’s model card.
TAKE IT FOR A SPIN
Start here.
Start with the source repository for evaluation tooling, or choose a model format below. The source README documents the Mac and Linux paths separately.
STATUS & REQUIREMENTS
Research artifacts derived from an upstream base model and dataset. The combined upstream license conditions apply; the model cards describe them.