Choose a workflow

Plan before downloading weights

Use the curated catalogue offline, or config/parameter metadata for another supported HF causal model. Metadata lookup does not load weights.

canifinetune doctor
canifinetune estimate --model Qwen/Qwen2.5-1.5B-Instruct --method qlora --gpu-vram-gb 16 --seq-len 2048 --offline
canifinetune recommend --model Qwen/Qwen2.5-0.5B-Instruct --gpu-vram-gb 16 --offline --top-k 3

Keep actual total VRAM in --gpu-vram-gb. On a busy device, also pass the current free budget with --available-vram-gb; overhead is still based on actual capacity. Every *_gb JSON value uses GiB. Inspect assumptions before acting on YES/NO.

Generate and execute a recipe

Install a qualified training stack, then create a fresh folder:

canifinetune recipe --model Qwen/Qwen2.5-0.5B-Instruct --revision 7ae557604adf67be50417f59c2c2f167def9a775 --method qlora --seq-len 256 --max-steps 2 --grad-accum 1 --output my-recipe
python my-recipe/train.py --config my-recipe/config.yaml
python my-recipe/eval_smoke.py --output-dir my-recipe/output --max-new-tokens 8

Training downloads weights when they are not cached. Check output/run.json: status must be success, update count must match, effective settings must match your intended workload, and reload must also succeed. Full saves model; LoRA/QLoRA saves adapter. Existing output directories are protected. Read the data contract before replacing sample records.

Measure a bounded workload

canifinetune bench --model Qwen/Qwen2.5-0.5B-Instruct --method qlora --seq-len 256 --optimizer adamw_torch --steps 3 --out-dir measurements
canifinetune report --benchmarks measurements --out report.md
canifinetune compare --benchmarks measurements --out compare.md

Bench loads weights and uses full-length synthetic labels. It can use more memory than the short recipe samples. Measurements cover loading through the declared updates, including first optimizer-state allocation. Compare the process reserved peak with estimated_process_reserved_gb, not the total planning budget including safety. Three updates do not demonstrate arbitrary long-run stability.

Fit or contribute evidence deliberately

calibrate fits eligible development observations; it is not independent validation. Compatibility checks include estimator version, configuration profile, model family and GPU capacity. Community/validation/failed observations are excluded. See evidence before exporting or fitting.