Most fine-tuning libraries force a choice: either drop down to distributed-training plumbing (FSDP, sharding, checkpoint juggling) or accept a rigid trainer that only does what its authors anticipated. This cookbook sits on a different layer. The Tinker API exposes just four primitives — forward_backward, optim_step, save_state, sample — and Thinking Machines Lab runs the cluster behind them. The recipes here show what you can build once that loop is yours but the infrastructure isn't your problem.
What Sets It Apart
- The abstraction boundary is unusually low: you own the optimization loop and reward logic, not the GPU topology. That means a custom RL algorithm is a few hundred readable lines, not a fork of a framework.
- Coverage spans the full post-training stack — chat SFT (Tulu3), math RL with verifiable rewards, sandboxed code RL, DPO, a three-stage RLHF pipeline (SFT, reward model, RL), on/off-policy distillation, retrieval tool use (a Search-R1 replication), and multi-agent self-play.
- A bundled benchmark harness scores trained models on ~12 evals (GSM8K, MMLU-Pro, IFEval, MBPP, C-Eval) with published reference numbers, so a recipe is verifiable rather than aspirational.
Who It's For
Great fit if you want to prototype or modify a post-training algorithm and care about the math more than the cluster — examples default to Llama-3.2-1B with LoRA, so iteration is cheap. Look elsewhere if you need full control over the training stack, on-prem weights, or a provider-agnostic setup: everything routes through the hosted Tinker service, which is the deliberate trade for not managing distributed training yourself.
