tinker-training-inference

1# Tinker: training and inference
2 
3Tinker gives you token-level control of hosted fine-tuning: you build the
4`Datum`s, you own the chat template, you hold the exact prompt and completion
5token ids on both sides. That is the reason to use it — a training run stays
6auditable, because the tokens the policy was updated on are tokens you can point
7at — and it is also why there is more to get right than with a hosted
8fine-tuning button. This skill is the map.
9 
10## Do this first, in any new Tinker project
11 
12**Create the pricing file.** Prices next to the training script get read before a
13550B RL run is launched; prices in a browser tab do not.
14 
15```bash
16python <skill>/scripts/tinker_pricing.py init --out <project dir>
17```
18 
19That writes `TINKER_PRICING.md` — the full catalogue, what each operation
20meters, and the pre-launch checklist. It is a **snapshot**: verify against the
21live catalogue at <https://tinker-docs.thinkingmachines.ai/tinker/models/>, and
22regenerate with `--force` when it moves. Then:
23 
24```bash
25export TINKER_API_KEY="..."          # console.tinker.ai
26uv pip install tinker
27```
28 
29## Where to go
30 
31| you want to | read |
32| --- | --- |
33| get set up, pick a model, avoid the context-window trap | [references/setup.md](references/setup.md) |
34| supervised fine-tuning | [references/training-sft.md](references/training-sft.md) |
35| RL / RFT on single-turn tasks | [references/training-rl.md](references/training-rl.md) |
36…