New Training — Benchgen

Select a base model and training dataset from HuggingFace, BenchGen, or your saved assets

Training Name

Saved to Knowledge API and reused as the default name when pushing the merged model to BenchGen.

Base Model

Dataset

Low-Rank Adaptation parameters

Rank (r)(4-64)8

464

Alpha(4-128)16

4128

Dropout(0.0-0.5)0.05

0.000.50

Learning Rate

NameQwen2.5-0.5B-Instruct - r8

MethodSFT

ModelQwen2.5-0.5B-Instruct

DatasetCapybara

GPU SelectionAuto

LoRA Rank8

Alpha16

Dropout0.05

Learn Rate1e-4

Epochs2

Batch Size4

Max Steps50

Rank 8-16 works well for most tasks

Alpha = 2xrank is the recommended ratio

50-100 steps for a quick validation run

Lower LR = more stable convergence

You'll be sent to login and brought right back to continue.