Capabilities
Fine-Tuning
Custom model training on user data
Read this when
- Training a model on customer data
Compatibility
| Provider | API | pi-go | Notes |
|---|---|---|---|
| Anthropic | ⚠️ Bedrock-only (Haiku, LoRA) | ❌ | |
| OpenAI Chat | ✅ SFT, vision FT, DPO, RFT | ❌ | |
| OpenAI Responses | — | — | uses base FT’d models |
| Google Gemini | ✅ supervised tuning (LoRA / full) on Gemini 2.5 Pro/Flash/Flash-lite via Vertex | ❌ | |
| Claude CLI | ❌ | — |
Provider Documentation
pi-go Gaps
- No fine-tuning helpers anywhere — typically out of scope for an SDK focused on inference.
- Fine-tuned model IDs do work today as plain
Model.IDstrings; no special handling required.