CLI Reference¶
The weather-extract command is the primary interface to the project.
weather-extract <command> [options]
Commands¶
info¶
Print the current project configuration as JSON.
weather-extract info
stages¶
Print a human-readable description of each pipeline stage.
weather-extract stages
ingest¶
Scan the sample directory and report how many image/transcription pairs exist.
weather-extract ingest
extract¶
Run a VLM over a single image and print the extracted JSON.
weather-extract extract [--model <name|id|path>] <image_path>
Flag |
Description |
|---|---|
|
Model preset ( |
evaluate¶
Evaluate a model over all paired images and print accuracy statistics.
weather-extract evaluate [--model <name>] [--limit N] [--tolerance F]
Flag |
Default |
Description |
|---|---|---|
|
|
Model or adapter path |
|
all |
Maximum images to evaluate |
|
|
Tolerance for cell match (inches) |
finetune¶
Fine-tune a base model on the paired dataset using LoRA.
weather-extract finetune [--model <name>] [--epochs N] [--output-dir PATH] \
[--eval-split F] [--lora-r N]
Flag |
Default |
Description |
|---|---|---|
|
|
Base model to fine-tune |
|
|
Training epochs |
|
|
Where to save the adapter |
|
|
Validation fraction |
|
|
LoRA rank |
visualize¶
Build a diagnostic figure (PNG) showing the image alongside the extracted table.
weather-extract visualize [--model <name>] [--output PATH] \
[--ground-truth PATH] <image_path>
Flag |
Default |
Description |
|---|---|---|
|
(none) |
Model or adapter for extraction |
|
|
Output path |
|
(sibling JSON) |
Ground-truth file |
Model presets¶
Preset |
HuggingFace ID |
|---|---|
|
|
|
|
|
|
|
|
Any full HuggingFace model ID or a local adapter directory can be passed to
--model in place of a preset name.
