Start with the result you want
- Generate small SVG drawings: start with the SVG exercise. You can inspect a generated sample alongside the training loss.
- Score a decision using existing features: use the Clef FP32 head workflow. Supply the approved package, feature vectors and labels, train only the head, then export, reload and score held-out examples. Decision-head workflow
- Experiment with a LoRA adapter: use the registered
decoder_lora_fp32_v1recipe. It accepts a bounded two-layer dense FP32 model, 2–64 fixed tokens, width up to 256 and vocabulary up to 4,096. Prepare separate tokenized training and validation examples and register the package before submitting a job. Package and job instructions
Can I train a model from the inference catalog?
Choose a recipe that explicitly names your input format, trainable weights and precision. A GGUF chat command only establishes inference for that artifact. General QLoRA or full-model fine-tuning of Inkling and the large chat models is outside the recipes listed here.
How do I know training helped?
Keep a few examples out of training. Save the starting outputs, train, reload the saved result and repeat the same examples. A completed job tells you the run finished; comparing those outputs tells you whether your task improved.
Train a tiny language model from scratch
Run the four-layer FP32 English fixture from the repository root. This reduced Qwen3-style example uses AdamW and checks restart and export.
make v8-training-workflow-fp32Inspect the held-out loss in the run report before trying your own data. FP32 training and notebook commands