Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

3.8. Diagnostics and Tracing

Without --verbose, LoRA training prints a single-line progress bar (pass N . loss . bar . % . ETA). Percent is loss progress from the pass-2 baseline toward the loss target, not pass/max-iters. Pass --verbose / -v for full [TRACE] output:

LineWhat it tells you
[TRACE] model type (chat template key) : tinyllamaWhether the template matches the model
[train-qa] iter=N loss=...Per-pass loss during training
[TRACE] inference model type: tinyllamaTemplate key at inference; must match training

If the template key at training and inference differ, the model will not recall trained facts. Rename the model file to include the architecture keyword (tinyllama, llama-3, mistral, phi3, qwen3) so ChatModelType.fromPath() picks the matching chat template. Qwen2/2.5 use ChatML; Qwen3 training uses the empty <think> block. LoRA training supports those dense architectures via LoraTrainingHandlerFactory; Gemma, Qwen3-MoE, and Qwen3.5 LoRA remain unsupported.

See also


<- 3.7 Test Mode  |  Table of Contents  |  4.1 Concepts ->