Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Troubleshooting

This page collects the errors you are most likely to hit and what they mean.

Dataset errors

A malformed dataset names the file and, for JSONL, the line.

SymptomLikely causeFix
unknown field ...A question carries a key outside the contract plus answerRemove extra keys
invalid typeA field has the wrong type (for example epochs = "three")Match the documented type
Empty or missing questionsA record without questionsEnsure questions is a non-empty object
choice rejectedNo criteria, or criteria emptyDeclare at least one option
score rejectedFewer than 2 or more than 10 levelsUse 2 to 10 ordered levels
answer not matchedThe answer is not a declared candidateUse a level name or option key
Item skippedPrompt longer than --max-sequence-lengthRaise the limit or shorten states

See Preparing datasets for the exact shapes.

Configuration errors

configuration file error in 'training.toml': unknown field `epocs`, expected one of ...
configuration file error in 'training.toml': invalid type: string "three", expected usize ...

Unknown keys and wrong types are rejected rather than ignored. A missing file is reported as an I/O error carrying the path. Remember the precedence: CLI flag > TOML key > default.

Model loading errors

SymptomMeaningFix
401 on downloadGated model without a tokenAccept the terms and set HF_TOKEN
MoE family rejectedmixtral, qwen3_moe, deepseek_v2/deepseek_v3Not supported; use a dense family
GGUF non-Qwen2 rejectedGGUF serving is Qwen2-onlyConvert to dense or use a Qwen2 GGUF
GPTQ/AWQ rejectedUnsupported quantizationConvert to FP8/FP4 or dense

API errors

Errors use the envelope {"error": {"message": "..."}}.

StatusSituation
422Malformed body or a question outside the contract
404Unknown model name
500Inference failure

Training misbehaves

SymptomLikely causeFix
Loss does not moveLearning rate too small, or label imbalanceRaise the LR; rebalance the dataset
Perfect train loss, poor servingOverfitting or leakageAdd data, lower rank, remove answers from states
Out of memory (full/from-scratch)Every parameter keeps F32 master + momentsUse LoRA/QLoRA or a smaller geometry
Adapter has no effectThe adapter was not merged before servingRun quantize --adapter-directory

Where to look next