Model
The base checkpoint. Smaller models cost less and finish sooner. This forge stops above 14B.
Cenedril small-model works · 14B and under
Bring a model.
Leave with an artifact.
Train, merge, and quantize from links. The Foundry inspects the cargo first, explains every conflict in plain words, then prepares a free-notebook run or a reproducible GPU Job billed only to your account.
The forge
One chart. Four bearings.
Work down the page. Yellow marks a warning. Red means stop and change course.
New here?
The short apprentice’s guide.
Dataset
Examples the model learns from. Quantizers such as GPTQ and AWQ also use examples to calibrate.
Training
SFT teaches examples. DPO and KTO align preferences. GRPO learns from verifiable rewards. Reward modeling builds a scorer.
Quantization
Weights use fewer bits. An importance matrix can preserve the weights your representative data activates most.
Output chart
Pick for the runtime that will use it.
FamilyBest useWatch for
bitsandbytesTransformers inference; broad architecture reachRuntime quantization, not a universal standalone file
GGUF · IQ/iMatrixActivation-aware local CPU/GPU inferenceNeeds representative text and architecture support
GPTQ / AWQFast GPU inference enginesNeeds calibration data; multimodal support varies
HQQ / torchao / QuantoSpecialized low-bit experimentsMatch the downstream runtime before forging
BF16 / FP16Portable merged weights and later conversionLarge output; not low-bit quantization