Skip to main content
← back to extensions

@swampadmin/llm-toolkit

v2026.02.14.0

End-to-end LLM fine-tuning pipeline with dataset preparation, distributed training, quantization, and automated evaluation.

Supported runtimes

Runtime Status
PyTorch Stable
JAX Beta
ONNX Export only

Repository

https://github.com/swamp-club/llm-toolkit

Platforms

linux-x86_64linux-aarch64

Labels

mlllmfine-tuningai

Install

$ swamp extension pull @swampadmin/llm-toolkit

Release Notes

New

  • Evaluation model with benchmark, compare, and report methods
  • Batch inference workflow — load model, run inference, generate report
  • Dataset augment method for synthetic data generation
  • Training quantize method for INT8/FP16 model compression

Breaking

  • training.export now requires a format argument (pytorch, onnx, safetensors)

@swampadmin/llm-toolkit/datasetv1.0.0dataset.ts
prepareprepare operation
ArgumentTypeDescription
namestringResource name
validatevalidate operation
ArgumentTypeDescription
namestringResource name
splitsplit operation
ArgumentTypeDescription
namestringResource name
augmentaugment operation
ArgumentTypeDescription
namestringResource name

Files

dataset.log(text/plain)— Operation audit log
dataset.json(application/json)— Structured output
@swampadmin/llm-toolkit/trainingv1.0.0training.ts
fine_tunefine tune operation
ArgumentTypeDescription
namestringResource name
resumeresume operation
ArgumentTypeDescription
namestringResource name
exportexport operation
ArgumentTypeDescription
namestringResource name
quantizequantize operation
ArgumentTypeDescription
namestringResource name

Resources

training.state(persistent)— Managed resource state
training.lock(ephemeral)— Concurrency lock

Files

training.log(text/plain)— Operation audit log
training.json(application/json)— Structured output
@swampadmin/llm-toolkit/evaluationv1.0.0evaluation.ts
benchmarkbenchmark operation
ArgumentTypeDescription
namestringResource name
comparecompare operation
ArgumentTypeDescription
namestringResource name
reportreport operation
ArgumentTypeDescription
namestringResource name

Files

evaluation.log(text/plain)— Operation audit log
evaluation.json(application/json)— Structured output

Train and Evaluatetrain-eval

Train and Evaluate workflow

train-eval-jobExecute Train and Evaluate
1.Prepare Dataset@swampadmin/llm-toolkit/dataset.prepare— Prepare Dataset step
2.Fine Tune@swampadmin/llm-toolkit/training.fine_tune— Fine Tune step
3.Evaluate@swampadmin/llm-toolkit/evaluation.benchmark— Evaluate step
Batch Inferencebatch-inference

Batch Inference workflow

batch-inference-jobExecute Batch Inference
1.Load Model@swampadmin/llm-toolkit/training.export— Load Model step
2.Run Inference@swampadmin/llm-toolkit/evaluation.benchmark— Run Inference step
3.Generate Report@swampadmin/llm-toolkit/evaluation.report— Generate Report step