ChartForge

Training appendix

Together for inference + FT · never OpenAI · FT paused

Together FT · launch paused · SOAP uses Together chat

Train on Together

Upload clinical dialogue→note JSONL, create a LoRA job (requires TOGETHER_FT_LAUNCH_ENABLED=1 + user). SOAP generate uses Together chat — never OpenAI cascade.

TOGETHER_API_KEY missingBase: …Data: clinical_dialogue_to_note.jsonlRunPod not configured
Open Together FT console ↗

Job status

Click Train on Together to upload JSONL and create a fine-tune job.

How to watch training in the Together UI

  1. Open https://api.together.ai/fine-tuning
  2. Select the job created from this panel.
  3. Use the Metrics tab for live train/eval loss curves.
  4. Lifecycle: pending → queued → running → uploading → completed.

Architecture appendix — not the product UI

RunPods training → managed serve

Real loop: CC-BY dialogue→note JSONL → SFT/LoRA (PEFT) → eval checkpoint → optional DPO. Together for inference + FT · never OpenAI — classic CAT local; flat = direct-prompt baseline; FT launch paused.

Local trainer (no RUNPOD_API_KEY)Train: Together Fine-Tuning (create locked / paused unless explicitly enabled)Serve: Fireworks / Together / BasetenData: clinical_dialogue_to_note.jsonl

Launch config

Dataset → train → eval → promote

  1. dataset
  2. train
  3. eval
  4. promote

Create a job to walk the training loop.

LoRA job YAML (RunPods)

Loading…

Train script

Loading…