Console

Fine Tuning as a Service

Preview a dataset, register it, choose a framework & PEFT technique, run real fine-tuning (torch / transformers / peft / TRL), track with MLflow, deploy, and prompt the adapted model.

1. Preview & register dataset

Datasets

2. Create fine-tune job

Default distilgpt2 is tuned for CPU demos (~minutes). Avoid sshleifer/tiny-gpt2 (gibberish). For ~7B use technique qlora on a CUDA host (Dockerfile.gpu) or set FTAAS_TRAIN_MODE=remote — see docs/gpu-training.md.

Jobs

JobModelFW / TechStatusProgressPipelineActions
job_9e91910cae4a distilgpt2 transformers / lora succeeded
Complete 		pl_335d76f397d4

Select a job to view logs

Click View log on a job (or the row) to stream logs and progress.

3. Deploy model

4. Prompt (UI / API)

Prompts use your train set: exact matches return the labeled answer; otherwise few-shot + the adapter generate. Ask something from the dataset (e.g. What is LoRA?). Very short trains (2 steps) will still hallucinate on novel questions.

Catalog

transformers trl verl llama-factory unsloth axolotl

Roadmap

  1. Phase 0 — Fine-Tuning & RL Templates (available)
  2. Phase 1 — Fine-Tuning UI (available)
  3. Phase 2 — Fine-Tune & Evaluate (available)
  4. Phase 3 — Resource Optimization (planned)
  5. Phase 4 — Sweeps & Optimization (planned)