Hands-on, end-to-end cookbooks for building with synthetic and de-identified data — each one a runnable notebook backed by a real experiment.
| Cookbook | Products | Run it |
|---|---|---|
Safely fine-tuning with sensitive data — de-identify clinical notes with consistent synthetic PII, fine-tune on the result, and measure that synthesis costs nothing (textual_deid_fireworks_sft.ipynb) |
Tonic Textual + Fireworks AI |  |
Zero-label fine-tuning with synthetic conversations — generate clinical conversations structure-first so extraction labels are correct by construction, then trace the accuracy-vs-data curve (fabricate_transcripts_fireworks_sft.ipynb) |
Tonic Fabricate + Fireworks AI | |
Teaching an AI agent your policies with RFT — reinforcement fine-tune an agent against a fully synthetic store whose APIs enforce a written policy, then show the tuned agent follows a revised policy document without retraining (fabricate_data_apis_fireworks_rft.ipynb) |
Tonic Fabricate + Fireworks AI | |
textual_deid_fireworks_sft.ipynb— the Textual × Fireworks cookbook: synthesis-based de-identification with per-record consistency, two-arm LoRA fine-tuning (synthetic vs. original data), and evaluation on real held-out notes.fabricate_transcripts_fireworks_sft.ipynb— the Fabricate × Fireworks cookbook: structure-first synthetic conversations with machine-verified labels (TonicAI/synthetic_clinical_conversations), a zero-shot capability ladder, and a three-tier fine-tuning learning curve.fabricate_data_apis_fireworks_rft.ipynb— the Fabricate × Fireworks RFT cookbook: a synthetic retail environment whose mock APIs enforce a written store policy, deterministic episode rewards (exact action match, no LLM judge), reinforcement fine-tuning of qwen3-32b (held-out policy application 44% → 67%), and policy-variation experiments (prompt swap + 8-policy grid) showing the policy stays a prompt-level control rather than being baked into the weights.
All notebooks run on a free Colab CPU runtime; the heavy lifting happens on Tonic and Fireworks infrastructure.
Each notebook states its own API-key and cost requirements up front.
Clinical examples use TonicAI/synthetic_clinical_notes (derived from Synthea patient records) and TonicAI/synthetic_clinical_conversations (generated structure-first with Tonic Fabricate). Both datasets are fully synthetic — no real patient data appears anywhere in these cookbooks. The retail-agent cookbook generates its entire environment (users, orders, products, scenarios) live in Fabricate — every record is synthetic. The frozen reference world from the published run is available as TonicAI/synthetic_retail_policy_environment.